
(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 5 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 (* -1.0 (* (floor w) dY.u)))
(t_2 (* -1.0 (* (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 = -1.0f * (floorf(w) * dY_46_u);
float t_2 = -1.0f * (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(Float32(-1.0) * Float32(floor(w) * dY_46_u)) t_2 = Float32(Float32(-1.0) * 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 = single(-1.0) * (floor(w) * dY_46_u); t_2 = single(-1.0) * (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 := -1 \cdot \left(\left\lfloor w\right\rfloor \cdot dY.u\right)\\
t_2 := -1 \cdot \left(\left\lfloor w\right\rfloor \cdot dX.u\right)\\
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}
Initial program 75.3%
Final simplification75.3%
(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)
(pow
(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)))
-0.5))
dX.v)
(* (/ 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) * powf(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))), -0.5f)) * dX_46_v;
} 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(-1.0) * Float32(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(Float32(floor(h) * (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))) ^ Float32(-0.5))) * dX_46_v); 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) * (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))) ^ single(-0.5))) * dX_46_v; 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 := -1 \cdot \left(\left\lfloor w\right\rfloor \cdot dY.u\right)\\
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(\left\lfloor h\right\rfloor \cdot {\left(\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)\right)}^{-0.5}\right) \cdot dX.v\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_7\right)}} \cdot t\_6\\
\end{array}
\end{array}
Initial program 75.3%
Applied rewrites75.2%
Final simplification75.2%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (log (floor h)) 2.0))
(t_2 (* -1.0 (* (floor w) dY.u)))
(t_3 (* -1.0 (* (floor w) dX.u)))
(t_4 (+ (* t_3 t_3) (* t_0 t_0)))
(t_5 (* (floor h) dY.v))
(t_6 (+ (* t_2 t_2) (* t_5 t_5)))
(t_7 (pow (floor w) 2.0)))
(if (>= t_4 t_6)
(*
(*
(floor h)
(pow
(fmax
(-
(* (* (+ (cosh t_1) (sinh t_1)) dX.v) dX.v)
(* (* -1.0 (* dX.u t_7)) dX.u))
(-
(* (* (pow (floor h) 2.0) dY.v) dY.v)
(* (* -1.0 (* dY.u t_7)) dY.u)))
-0.5))
dX.v)
(* (/ 1.0 (sqrt (fmax t_4 t_6))) t_5))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = logf(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 = (t_3 * t_3) + (t_0 * t_0);
float t_5 = floorf(h) * dY_46_v;
float t_6 = (t_2 * t_2) + (t_5 * t_5);
float t_7 = powf(floorf(w), 2.0f);
float tmp;
if (t_4 >= t_6) {
tmp = (floorf(h) * powf(fmaxf(((((coshf(t_1) + sinhf(t_1)) * dX_46_v) * dX_46_v) - ((-1.0f * (dX_46_u * t_7)) * dX_46_u)), (((powf(floorf(h), 2.0f) * dY_46_v) * dY_46_v) - ((-1.0f * (dY_46_u * t_7)) * dY_46_u))), -0.5f)) * dX_46_v;
} else {
tmp = (1.0f / sqrtf(fmaxf(t_4, t_6))) * t_5;
}
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(log(floor(h)) * Float32(2.0)) t_2 = Float32(Float32(-1.0) * Float32(floor(w) * dY_46_u)) t_3 = Float32(Float32(-1.0) * Float32(floor(w) * dX_46_u)) t_4 = Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)) t_5 = Float32(floor(h) * dY_46_v) t_6 = Float32(Float32(t_2 * t_2) + Float32(t_5 * t_5)) t_7 = floor(w) ^ Float32(2.0) tmp = Float32(0.0) if (t_4 >= t_6) tmp = Float32(Float32(floor(h) * (fmax(Float32(Float32(Float32(Float32(cosh(t_1) + sinh(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((floor(h) ^ Float32(2.0)) * dY_46_v) * dY_46_v) - Float32(Float32(Float32(-1.0) * Float32(dY_46_u * t_7)) * dY_46_u))) ^ Float32(-0.5))) * dX_46_v); else tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(t_4, t_6))) * t_5); 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 = log(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 = (t_3 * t_3) + (t_0 * t_0); t_5 = floor(h) * dY_46_v; t_6 = (t_2 * t_2) + (t_5 * t_5); t_7 = floor(w) ^ single(2.0); tmp = single(0.0); if (t_4 >= t_6) tmp = (floor(h) * (max(((((cosh(t_1) + sinh(t_1)) * dX_46_v) * dX_46_v) - ((single(-1.0) * (dX_46_u * t_7)) * dX_46_u)), ((((floor(h) ^ single(2.0)) * dY_46_v) * dY_46_v) - ((single(-1.0) * (dY_46_u * t_7)) * dY_46_u))) ^ single(-0.5))) * dX_46_v; else tmp = (single(1.0) / sqrt(max(t_4, t_6))) * t_5; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \log \left(\left\lfloor h\right\rfloor \right) \cdot 2\\
t_2 := -1 \cdot \left(\left\lfloor w\right\rfloor \cdot dY.u\right)\\
t_3 := -1 \cdot \left(\left\lfloor w\right\rfloor \cdot dX.u\right)\\
t_4 := t\_3 \cdot t\_3 + t\_0 \cdot t\_0\\
t_5 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_6 := t\_2 \cdot t\_2 + t\_5 \cdot t\_5\\
t_7 := {\left(\left\lfloor w\right\rfloor \right)}^{2}\\
\mathbf{if}\;t\_4 \geq t\_6:\\
\;\;\;\;\left(\left\lfloor h\right\rfloor \cdot {\left(\mathsf{max}\left(\left(\left(\cosh t\_1 + \sinh t\_1\right) \cdot dX.v\right) \cdot dX.v - \left(-1 \cdot \left(dX.u \cdot t\_7\right)\right) \cdot dX.u, \left({\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot dY.v\right) \cdot dY.v - \left(-1 \cdot \left(dY.u \cdot t\_7\right)\right) \cdot dY.u\right)\right)}^{-0.5}\right) \cdot dX.v\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_4, t\_6\right)}} \cdot t\_5\\
\end{array}
\end{array}
Initial program 75.3%
Applied rewrites75.2%
lift-floor.f32N/A
lift-pow.f32N/A
pow-to-expN/A
sinh-+-cosh-revN/A
lower-+.f32N/A
lower-cosh.f32N/A
lower-*.f32N/A
lower-log.f32N/A
lift-floor.f32N/A
lower-sinh.f32N/A
lower-*.f32N/A
lower-log.f32N/A
lift-floor.f3274.9
Applied rewrites74.9%
Final simplification74.9%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* -1.0 (* (floor w) dY.u)))
(t_1 (pow (floor h) 2.0))
(t_2 (* (* t_1 dY.v) dY.v))
(t_3 (* (* t_1 dX.v) dX.v))
(t_4 (* (floor h) dX.v))
(t_5 (pow (floor w) 2.0))
(t_6 (* -1.0 (* (floor w) dX.u)))
(t_7 (+ (* t_6 t_6) (* t_4 t_4)))
(t_8 (* (floor h) dY.v))
(t_9 (* t_5 dY.u))
(t_10 (* t_9 dY.u))
(t_11 (+ (* t_0 t_0) (* t_8 t_8))))
(if (or (<= dX.v -0.5) (not (<= dX.v 40000000000.0)))
(if (>= t_7 t_10)
(/ (* dX.v (floor h)) (sqrt (fmax t_3 t_10)))
(* (/ 1.0 (sqrt (fmax t_7 t_11))) t_8))
(if (>= t_7 t_11)
(*
(exp
(*
(log
(sqrt
(fmax
(* (* (* t_5 dX.u) -1.0) (* -1.0 dX.u))
(- t_2 (* (* -1.0 t_9) dY.u)))))
-1.0))
t_4)
(/
(* dY.v (floor h))
(sqrt
(fmax
(- t_3 (* (* -1.0 (* dX.u t_5)) dX.u))
(- t_2 (* (* -1.0 (* dY.u t_5)) 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 = -1.0f * (floorf(w) * dY_46_u);
float t_1 = powf(floorf(h), 2.0f);
float t_2 = (t_1 * dY_46_v) * dY_46_v;
float t_3 = (t_1 * dX_46_v) * dX_46_v;
float t_4 = floorf(h) * dX_46_v;
float t_5 = powf(floorf(w), 2.0f);
float t_6 = -1.0f * (floorf(w) * dX_46_u);
float t_7 = (t_6 * t_6) + (t_4 * t_4);
float t_8 = floorf(h) * dY_46_v;
float t_9 = t_5 * dY_46_u;
float t_10 = t_9 * dY_46_u;
float t_11 = (t_0 * t_0) + (t_8 * t_8);
float tmp_1;
if ((dX_46_v <= -0.5f) || !(dX_46_v <= 40000000000.0f)) {
float tmp_2;
if (t_7 >= t_10) {
tmp_2 = (dX_46_v * floorf(h)) / sqrtf(fmaxf(t_3, t_10));
} else {
tmp_2 = (1.0f / sqrtf(fmaxf(t_7, t_11))) * t_8;
}
tmp_1 = tmp_2;
} else if (t_7 >= t_11) {
tmp_1 = expf((logf(sqrtf(fmaxf((((t_5 * dX_46_u) * -1.0f) * (-1.0f * dX_46_u)), (t_2 - ((-1.0f * t_9) * dY_46_u))))) * -1.0f)) * t_4;
} else {
tmp_1 = (dY_46_v * floorf(h)) / sqrtf(fmaxf((t_3 - ((-1.0f * (dX_46_u * t_5)) * dX_46_u)), (t_2 - ((-1.0f * (dY_46_u * t_5)) * dY_46_u))));
}
return tmp_1;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(Float32(-1.0) * Float32(floor(w) * dY_46_u)) t_1 = floor(h) ^ Float32(2.0) t_2 = Float32(Float32(t_1 * dY_46_v) * dY_46_v) t_3 = Float32(Float32(t_1 * dX_46_v) * dX_46_v) t_4 = Float32(floor(h) * dX_46_v) t_5 = floor(w) ^ Float32(2.0) t_6 = Float32(Float32(-1.0) * Float32(floor(w) * dX_46_u)) t_7 = Float32(Float32(t_6 * t_6) + Float32(t_4 * t_4)) t_8 = Float32(floor(h) * dY_46_v) t_9 = Float32(t_5 * dY_46_u) t_10 = Float32(t_9 * dY_46_u) t_11 = Float32(Float32(t_0 * t_0) + Float32(t_8 * t_8)) tmp_1 = Float32(0.0) if ((dX_46_v <= Float32(-0.5)) || !(dX_46_v <= Float32(40000000000.0))) tmp_2 = Float32(0.0) if (t_7 >= t_10) tmp_2 = Float32(Float32(dX_46_v * floor(h)) / sqrt(fmax(t_3, t_10))); else tmp_2 = Float32(Float32(Float32(1.0) / sqrt(fmax(t_7, t_11))) * t_8); end tmp_1 = tmp_2; elseif (t_7 >= t_11) tmp_1 = Float32(exp(Float32(log(sqrt(fmax(Float32(Float32(Float32(t_5 * dX_46_u) * Float32(-1.0)) * Float32(Float32(-1.0) * dX_46_u)), Float32(t_2 - Float32(Float32(Float32(-1.0) * t_9) * dY_46_u))))) * Float32(-1.0))) * t_4); else tmp_1 = Float32(Float32(dY_46_v * floor(h)) / sqrt(fmax(Float32(t_3 - Float32(Float32(Float32(-1.0) * Float32(dX_46_u * t_5)) * dX_46_u)), Float32(t_2 - Float32(Float32(Float32(-1.0) * Float32(dY_46_u * t_5)) * dY_46_u))))); 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 = single(-1.0) * (floor(w) * dY_46_u); t_1 = floor(h) ^ single(2.0); t_2 = (t_1 * dY_46_v) * dY_46_v; t_3 = (t_1 * dX_46_v) * dX_46_v; t_4 = floor(h) * dX_46_v; t_5 = floor(w) ^ single(2.0); t_6 = single(-1.0) * (floor(w) * dX_46_u); t_7 = (t_6 * t_6) + (t_4 * t_4); t_8 = floor(h) * dY_46_v; t_9 = t_5 * dY_46_u; t_10 = t_9 * dY_46_u; t_11 = (t_0 * t_0) + (t_8 * t_8); tmp_2 = single(0.0); if ((dX_46_v <= single(-0.5)) || ~((dX_46_v <= single(40000000000.0)))) tmp_3 = single(0.0); if (t_7 >= t_10) tmp_3 = (dX_46_v * floor(h)) / sqrt(max(t_3, t_10)); else tmp_3 = (single(1.0) / sqrt(max(t_7, t_11))) * t_8; end tmp_2 = tmp_3; elseif (t_7 >= t_11) tmp_2 = exp((log(sqrt(max((((t_5 * dX_46_u) * single(-1.0)) * (single(-1.0) * dX_46_u)), (t_2 - ((single(-1.0) * t_9) * dY_46_u))))) * single(-1.0))) * t_4; else tmp_2 = (dY_46_v * floor(h)) / sqrt(max((t_3 - ((single(-1.0) * (dX_46_u * t_5)) * dX_46_u)), (t_2 - ((single(-1.0) * (dY_46_u * t_5)) * dY_46_u)))); end tmp_4 = tmp_2; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := -1 \cdot \left(\left\lfloor w\right\rfloor \cdot dY.u\right)\\
t_1 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
t_2 := \left(t\_1 \cdot dY.v\right) \cdot dY.v\\
t_3 := \left(t\_1 \cdot dX.v\right) \cdot dX.v\\
t_4 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_5 := {\left(\left\lfloor w\right\rfloor \right)}^{2}\\
t_6 := -1 \cdot \left(\left\lfloor w\right\rfloor \cdot dX.u\right)\\
t_7 := t\_6 \cdot t\_6 + t\_4 \cdot t\_4\\
t_8 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_9 := t\_5 \cdot dY.u\\
t_10 := t\_9 \cdot dY.u\\
t_11 := t\_0 \cdot t\_0 + t\_8 \cdot t\_8\\
\mathbf{if}\;dX.v \leq -0.5 \lor \neg \left(dX.v \leq 40000000000\right):\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_7 \geq t\_10:\\
\;\;\;\;\frac{dX.v \cdot \left\lfloor h\right\rfloor }{\sqrt{\mathsf{max}\left(t\_3, t\_10\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_7, t\_11\right)}} \cdot t\_8\\
\end{array}\\
\mathbf{elif}\;t\_7 \geq t\_11:\\
\;\;\;\;e^{\log \left(\sqrt{\mathsf{max}\left(\left(\left(t\_5 \cdot dX.u\right) \cdot -1\right) \cdot \left(-1 \cdot dX.u\right), t\_2 - \left(-1 \cdot t\_9\right) \cdot dY.u\right)}\right) \cdot -1} \cdot t\_4\\
\mathbf{else}:\\
\;\;\;\;\frac{dY.v \cdot \left\lfloor h\right\rfloor }{\sqrt{\mathsf{max}\left(t\_3 - \left(-1 \cdot \left(dX.u \cdot t\_5\right)\right) \cdot dX.u, t\_2 - \left(-1 \cdot \left(dY.u \cdot t\_5\right)\right) \cdot dY.u\right)}}\\
\end{array}
\end{array}
if dX.v < -0.5 or 4e10 < dX.v Initial program 63.8%
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
lift-floor.f3258.3
Applied rewrites58.3%
Taylor expanded in dY.u around inf
Applied rewrites58.4%
Taylor expanded in dY.u around inf
*-commutativeN/A
unpow2N/A
associate-*l*N/A
lift-pow.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3258.3
Applied rewrites58.3%
Applied rewrites58.4%
if -0.5 < dX.v < 4e10Initial program 81.5%
Applied rewrites81.5%
Taylor expanded in dX.u around inf
Applied rewrites77.1%
lift-floor.f32N/A
lift-pow.f32N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f32N/A
lift-floor.f3276.6
Applied rewrites76.6%
Applied rewrites75.3%
Final simplification69.3%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* -1.0 (* (floor w) dY.u)))
(t_2 (* (floor h) dY.v))
(t_3 (pow (floor w) 2.0))
(t_4 (log t_3))
(t_5 (cosh t_4))
(t_6 (* -1.0 (* (floor w) dX.u)))
(t_7 (+ (* t_6 t_6) (* t_0 t_0)))
(t_8 (sinh t_4)))
(if (>= t_7 (* (* t_3 dY.u) dY.u))
(*
(/
1.0
(sqrt
(fmax
(* (* (pow (floor h) 2.0) dX.v) dX.v)
(*
(* (/ (- (* t_5 t_5) (* t_8 t_8)) (- t_5 t_8)) (* -1.0 dY.u))
(* -1.0 dY.u)))))
t_0)
(* (/ 1.0 (sqrt (fmax t_7 (+ (* t_1 t_1) (* t_2 t_2))))) t_2))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = -1.0f * (floorf(w) * dY_46_u);
float t_2 = floorf(h) * dY_46_v;
float t_3 = powf(floorf(w), 2.0f);
float t_4 = logf(t_3);
float t_5 = coshf(t_4);
float t_6 = -1.0f * (floorf(w) * dX_46_u);
float t_7 = (t_6 * t_6) + (t_0 * t_0);
float t_8 = sinhf(t_4);
float tmp;
if (t_7 >= ((t_3 * dY_46_u) * dY_46_u)) {
tmp = (1.0f / sqrtf(fmaxf(((powf(floorf(h), 2.0f) * dX_46_v) * dX_46_v), (((((t_5 * t_5) - (t_8 * t_8)) / (t_5 - t_8)) * (-1.0f * dY_46_u)) * (-1.0f * dY_46_u))))) * t_0;
} else {
tmp = (1.0f / sqrtf(fmaxf(t_7, ((t_1 * t_1) + (t_2 * t_2))))) * t_2;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(Float32(-1.0) * Float32(floor(w) * dY_46_u)) t_2 = Float32(floor(h) * dY_46_v) t_3 = floor(w) ^ Float32(2.0) t_4 = log(t_3) t_5 = cosh(t_4) t_6 = Float32(Float32(-1.0) * Float32(floor(w) * dX_46_u)) t_7 = Float32(Float32(t_6 * t_6) + Float32(t_0 * t_0)) t_8 = sinh(t_4) tmp = Float32(0.0) if (t_7 >= Float32(Float32(t_3 * dY_46_u) * dY_46_u)) tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(Float32(Float32((floor(h) ^ Float32(2.0)) * dX_46_v) * dX_46_v), Float32(Float32(Float32(Float32(Float32(t_5 * t_5) - Float32(t_8 * t_8)) / Float32(t_5 - t_8)) * Float32(Float32(-1.0) * dY_46_u)) * Float32(Float32(-1.0) * dY_46_u))))) * t_0); else tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(t_7, Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2))))) * t_2); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = single(-1.0) * (floor(w) * dY_46_u); t_2 = floor(h) * dY_46_v; t_3 = floor(w) ^ single(2.0); t_4 = log(t_3); t_5 = cosh(t_4); t_6 = single(-1.0) * (floor(w) * dX_46_u); t_7 = (t_6 * t_6) + (t_0 * t_0); t_8 = sinh(t_4); tmp = single(0.0); if (t_7 >= ((t_3 * dY_46_u) * dY_46_u)) tmp = (single(1.0) / sqrt(max((((floor(h) ^ single(2.0)) * dX_46_v) * dX_46_v), (((((t_5 * t_5) - (t_8 * t_8)) / (t_5 - t_8)) * (single(-1.0) * dY_46_u)) * (single(-1.0) * dY_46_u))))) * t_0; else tmp = (single(1.0) / sqrt(max(t_7, ((t_1 * t_1) + (t_2 * t_2))))) * t_2; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := -1 \cdot \left(\left\lfloor w\right\rfloor \cdot dY.u\right)\\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_3 := {\left(\left\lfloor w\right\rfloor \right)}^{2}\\
t_4 := \log t\_3\\
t_5 := \cosh t\_4\\
t_6 := -1 \cdot \left(\left\lfloor w\right\rfloor \cdot dX.u\right)\\
t_7 := t\_6 \cdot t\_6 + t\_0 \cdot t\_0\\
t_8 := \sinh t\_4\\
\mathbf{if}\;t\_7 \geq \left(t\_3 \cdot dY.u\right) \cdot dY.u:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(\left({\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot dX.v\right) \cdot dX.v, \left(\frac{t\_5 \cdot t\_5 - t\_8 \cdot t\_8}{t\_5 - t\_8} \cdot \left(-1 \cdot dY.u\right)\right) \cdot \left(-1 \cdot dY.u\right)\right)}} \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_7, t\_1 \cdot t\_1 + t\_2 \cdot t\_2\right)}} \cdot t\_2\\
\end{array}
\end{array}
Initial program 75.3%
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
lift-floor.f3257.9
Applied rewrites57.9%
Taylor expanded in dY.u around inf
Applied rewrites57.5%
Taylor expanded in dY.u around inf
*-commutativeN/A
unpow2N/A
associate-*l*N/A
lift-pow.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3250.9
Applied rewrites50.9%
lift-floor.f32N/A
lift-pow.f32N/A
pow-to-expN/A
sinh-+-cosh-revN/A
flip-+N/A
lower-/.f32N/A
Applied rewrites53.3%
Final simplification53.3%
herbie shell --seed 2025065
(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))))