
(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}
Herbie found 9 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 (* (* dX.v (floor h)) dX.v))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor w) dX.u))
(t_3 (* (floor h) dX.v))
(t_4 (+ (* t_2 t_2) (* t_3 t_3)))
(t_5 (* (floor h) dY.v))
(t_6 (+ (* t_1 t_1) (* t_5 t_5)))
(t_7 (* (floor w) (floor w)))
(t_8 (* (* dX.u dX.u) t_7))
(t_9 (/ 1.0 (sqrt (fmax t_4 t_6))))
(t_10 (if (>= t_4 t_6) (* t_9 t_3) (* t_9 t_5)))
(t_11 (* (floor h) (floor h)))
(t_12 (fma (* t_7 dX.u) dX.u (* t_11 (* dX.v dX.v))))
(t_13 (fma (* t_7 dY.u) dY.u (* (* dY.v dY.v) t_11)))
(t_14 (/ t_3 (sqrt (fmax t_12 t_13))))
(t_15
(if (>= (* t_0 (floor h)) t_13)
t_14
(*
(/
dY.v
(sqrt
(fmax
(fma t_0 (floor h) t_8)
(fma
(* t_1 dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h))))))
(floor h)))))
(if (<= t_10 -1.9999999494757503e-5)
t_15
(if (<= t_10 5.000000058430487e-8)
(if (>= t_8 t_13) t_14 (/ t_5 (sqrt (fmax t_12 (* (* dY.u dY.u) t_7)))))
t_15))))
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_v * floorf(h)) * dX_46_v;
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(w) * dX_46_u;
float t_3 = floorf(h) * dX_46_v;
float t_4 = (t_2 * t_2) + (t_3 * t_3);
float t_5 = floorf(h) * dY_46_v;
float t_6 = (t_1 * t_1) + (t_5 * t_5);
float t_7 = floorf(w) * floorf(w);
float t_8 = (dX_46_u * dX_46_u) * t_7;
float t_9 = 1.0f / sqrtf(fmaxf(t_4, t_6));
float tmp;
if (t_4 >= t_6) {
tmp = t_9 * t_3;
} else {
tmp = t_9 * t_5;
}
float t_10 = tmp;
float t_11 = floorf(h) * floorf(h);
float t_12 = fmaf((t_7 * dX_46_u), dX_46_u, (t_11 * (dX_46_v * dX_46_v)));
float t_13 = fmaf((t_7 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_11));
float t_14 = t_3 / sqrtf(fmaxf(t_12, t_13));
float tmp_1;
if ((t_0 * floorf(h)) >= t_13) {
tmp_1 = t_14;
} else {
tmp_1 = (dY_46_v / sqrtf(fmaxf(fmaf(t_0, floorf(h), t_8), fmaf((t_1 * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)))))) * floorf(h);
}
float t_15 = tmp_1;
float tmp_2;
if (t_10 <= -1.9999999494757503e-5f) {
tmp_2 = t_15;
} else if (t_10 <= 5.000000058430487e-8f) {
float tmp_3;
if (t_8 >= t_13) {
tmp_3 = t_14;
} else {
tmp_3 = t_5 / sqrtf(fmaxf(t_12, ((dY_46_u * dY_46_u) * t_7)));
}
tmp_2 = tmp_3;
} else {
tmp_2 = t_15;
}
return tmp_2;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(Float32(dX_46_v * floor(h)) * dX_46_v) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(floor(h) * dX_46_v) t_4 = Float32(Float32(t_2 * t_2) + Float32(t_3 * t_3)) t_5 = Float32(floor(h) * dY_46_v) t_6 = Float32(Float32(t_1 * t_1) + Float32(t_5 * t_5)) t_7 = Float32(floor(w) * floor(w)) t_8 = Float32(Float32(dX_46_u * dX_46_u) * t_7) t_9 = Float32(Float32(1.0) / sqrt(fmax(t_4, t_6))) tmp = Float32(0.0) if (t_4 >= t_6) tmp = Float32(t_9 * t_3); else tmp = Float32(t_9 * t_5); end t_10 = tmp t_11 = Float32(floor(h) * floor(h)) t_12 = fma(Float32(t_7 * dX_46_u), dX_46_u, Float32(t_11 * Float32(dX_46_v * dX_46_v))) t_13 = fma(Float32(t_7 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_11)) t_14 = Float32(t_3 / sqrt(fmax(t_12, t_13))) tmp_1 = Float32(0.0) if (Float32(t_0 * floor(h)) >= t_13) tmp_1 = t_14; else tmp_1 = Float32(Float32(dY_46_v / sqrt(fmax(fma(t_0, floor(h), t_8), fma(Float32(t_1 * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h)))))) * floor(h)); end t_15 = tmp_1 tmp_2 = Float32(0.0) if (t_10 <= Float32(-1.9999999494757503e-5)) tmp_2 = t_15; elseif (t_10 <= Float32(5.000000058430487e-8)) tmp_3 = Float32(0.0) if (t_8 >= t_13) tmp_3 = t_14; else tmp_3 = Float32(t_5 / sqrt(fmax(t_12, Float32(Float32(dY_46_u * dY_46_u) * t_7)))); end tmp_2 = tmp_3; else tmp_2 = t_15; end return tmp_2 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left(dX.v \cdot \left\lfloor h\right\rfloor \right) \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 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_4 := t\_2 \cdot t\_2 + t\_3 \cdot t\_3\\
t_5 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_6 := t\_1 \cdot t\_1 + t\_5 \cdot t\_5\\
t_7 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_8 := \left(dX.u \cdot dX.u\right) \cdot t\_7\\
t_9 := \frac{1}{\sqrt{\mathsf{max}\left(t\_4, t\_6\right)}}\\
t_10 := \begin{array}{l}
\mathbf{if}\;t\_4 \geq t\_6:\\
\;\;\;\;t\_9 \cdot t\_3\\
\mathbf{else}:\\
\;\;\;\;t\_9 \cdot t\_5\\
\end{array}\\
t_11 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_12 := \mathsf{fma}\left(t\_7 \cdot dX.u, dX.u, t\_11 \cdot \left(dX.v \cdot dX.v\right)\right)\\
t_13 := \mathsf{fma}\left(t\_7 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_11\right)\\
t_14 := \frac{t\_3}{\sqrt{\mathsf{max}\left(t\_12, t\_13\right)}}\\
t_15 := \begin{array}{l}
\mathbf{if}\;t\_0 \cdot \left\lfloor h\right\rfloor \geq t\_13:\\
\;\;\;\;t\_14\\
\mathbf{else}:\\
\;\;\;\;\frac{dY.v}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, \left\lfloor h\right\rfloor , t\_8\right), \mathsf{fma}\left(t\_1 \cdot dY.u, \left\lfloor w\right\rfloor , \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)}} \cdot \left\lfloor h\right\rfloor \\
\end{array}\\
\mathbf{if}\;t\_10 \leq -1.9999999494757503 \cdot 10^{-5}:\\
\;\;\;\;t\_15\\
\mathbf{elif}\;t\_10 \leq 5.000000058430487 \cdot 10^{-8}:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_8 \geq t\_13:\\
\;\;\;\;t\_14\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_5}{\sqrt{\mathsf{max}\left(t\_12, \left(dY.u \cdot dY.u\right) \cdot t\_7\right)}}\\
\end{array}\\
\mathbf{else}:\\
\;\;\;\;t\_15\\
\end{array}
\end{array}
if (if (>=.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dX.v)) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dY.v))) < -1.99999995e-5 or 5.00000006e-8 < (if (>=.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dX.v)) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dY.v))) Initial program 99.3%
Applied rewrites99.2%
Applied rewrites99.1%
Taylor expanded in dX.u around 0
Applied rewrites97.3%
if -1.99999995e-5 < (if (>=.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dX.v)) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dY.v))) < 5.00000006e-8Initial program 56.2%
Applied rewrites56.3%
Taylor expanded in dX.u around inf
Applied rewrites56.3%
Taylor expanded in dY.u around inf
Applied rewrites57.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* (floor w) (floor w)))
(t_2 (fma (* t_1 dY.u) dY.u (* (* dY.v dY.v) t_0)))
(t_3 (fma (* t_1 dX.u) dX.u (* t_0 (* dX.v dX.v)))))
(if (>= t_3 t_2)
(/ (* (floor h) dX.v) (sqrt (fmax t_3 t_2)))
(*
(/
dY.v
(sqrt
(fmax
(fma (* (* dX.v (floor h)) dX.v) (floor h) (* (* dX.u dX.u) t_1))
(fma
(* (* (floor w) dY.u) dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h))))))
(floor h)))))
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) * floorf(h);
float t_1 = floorf(w) * floorf(w);
float t_2 = fmaf((t_1 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_0));
float t_3 = fmaf((t_1 * dX_46_u), dX_46_u, (t_0 * (dX_46_v * dX_46_v)));
float tmp;
if (t_3 >= t_2) {
tmp = (floorf(h) * dX_46_v) / sqrtf(fmaxf(t_3, t_2));
} else {
tmp = (dY_46_v / sqrtf(fmaxf(fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), ((dX_46_u * dX_46_u) * t_1)), fmaf(((floorf(w) * dY_46_u) * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)))))) * floorf(h);
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * floor(h)) t_1 = Float32(floor(w) * floor(w)) t_2 = fma(Float32(t_1 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_0)) t_3 = fma(Float32(t_1 * dX_46_u), dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))) tmp = Float32(0.0) if (t_3 >= t_2) tmp = Float32(Float32(floor(h) * dX_46_v) / sqrt(fmax(t_3, t_2))); else tmp = Float32(Float32(dY_46_v / sqrt(fmax(fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(dX_46_u * dX_46_u) * t_1)), fma(Float32(Float32(floor(w) * dY_46_u) * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h)))))) * floor(h)); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_2 := \mathsf{fma}\left(t\_1 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_0\right)\\
t_3 := \mathsf{fma}\left(t\_1 \cdot dX.u, dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right)\\
\mathbf{if}\;t\_3 \geq t\_2:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dX.v}{\sqrt{\mathsf{max}\left(t\_3, t\_2\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{dY.v}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, \left\lfloor h\right\rfloor , \left(dX.u \cdot dX.u\right) \cdot t\_1\right), \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot dY.u\right) \cdot dY.u, \left\lfloor w\right\rfloor , \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)}} \cdot \left\lfloor h\right\rfloor \\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.1%
Applied rewrites76.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* (floor w) (floor w)))
(t_2 (fma (* t_1 dY.u) dY.u (* (* dY.v dY.v) t_0)))
(t_3 (fma (* t_1 dX.u) dX.u (* t_0 (* dX.v dX.v)))))
(if (>= t_3 t_2)
(*
(/
dX.v
(sqrt
(fmax
(fma (* (* dX.v (floor h)) dX.v) (floor h) (* (* dX.u dX.u) t_1))
(fma
(* (* (floor w) dY.u) dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h))))))
(floor h))
(/ (* (floor h) dY.v) (sqrt (fmax t_3 t_2))))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * floorf(h);
float t_1 = floorf(w) * floorf(w);
float t_2 = fmaf((t_1 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_0));
float t_3 = fmaf((t_1 * dX_46_u), dX_46_u, (t_0 * (dX_46_v * dX_46_v)));
float tmp;
if (t_3 >= t_2) {
tmp = (dX_46_v / sqrtf(fmaxf(fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), ((dX_46_u * dX_46_u) * t_1)), fmaf(((floorf(w) * dY_46_u) * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)))))) * floorf(h);
} else {
tmp = (floorf(h) * dY_46_v) / sqrtf(fmaxf(t_3, t_2));
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * floor(h)) t_1 = Float32(floor(w) * floor(w)) t_2 = fma(Float32(t_1 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_0)) t_3 = fma(Float32(t_1 * dX_46_u), dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))) tmp = Float32(0.0) if (t_3 >= t_2) tmp = Float32(Float32(dX_46_v / sqrt(fmax(fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(dX_46_u * dX_46_u) * t_1)), fma(Float32(Float32(floor(w) * dY_46_u) * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h)))))) * floor(h)); else tmp = Float32(Float32(floor(h) * dY_46_v) / sqrt(fmax(t_3, t_2))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_2 := \mathsf{fma}\left(t\_1 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_0\right)\\
t_3 := \mathsf{fma}\left(t\_1 \cdot dX.u, dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right)\\
\mathbf{if}\;t\_3 \geq t\_2:\\
\;\;\;\;\frac{dX.v}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, \left\lfloor h\right\rfloor , \left(dX.u \cdot dX.u\right) \cdot t\_1\right), \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot dY.u\right) \cdot dY.u, \left\lfloor w\right\rfloor , \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)}} \cdot \left\lfloor h\right\rfloor \\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dY.v}{\sqrt{\mathsf{max}\left(t\_3, t\_2\right)}}\\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.1%
Applied rewrites76.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* t_0 dY.u))
(t_2 (* (floor h) dX.v))
(t_3 (* (floor h) dY.v))
(t_4 (* dY.v (floor h)))
(t_5 (* dX.v (floor h)))
(t_6
(fma
(* t_5 dX.v)
(floor h)
(* (* dX.u dX.u) (* (floor w) (floor w)))))
(t_7 (sqrt (fmax t_6 (fma t_1 (floor w) (* (* t_4 dY.v) (floor h))))))
(t_8 (* (floor w) dX.u))
(t_9 (+ (* t_8 t_8) (* t_2 t_2)))
(t_10 (/ 1.0 (sqrt (fmax t_9 (+ (* t_0 t_0) (* t_3 t_3)))))))
(if (<= dY.u 8000000.0)
(if (>= t_9 (* (* dY.v dY.v) (* (floor h) (floor h))))
(* t_10 t_2)
(* t_10 t_3))
(if (>= t_6 (* t_1 (floor w))) (/ t_5 t_7) (/ t_4 t_7)))))
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 = t_0 * dY_46_u;
float t_2 = floorf(h) * dX_46_v;
float t_3 = floorf(h) * dY_46_v;
float t_4 = dY_46_v * floorf(h);
float t_5 = dX_46_v * floorf(h);
float t_6 = fmaf((t_5 * dX_46_v), floorf(h), ((dX_46_u * dX_46_u) * (floorf(w) * floorf(w))));
float t_7 = sqrtf(fmaxf(t_6, fmaf(t_1, floorf(w), ((t_4 * dY_46_v) * floorf(h)))));
float t_8 = floorf(w) * dX_46_u;
float t_9 = (t_8 * t_8) + (t_2 * t_2);
float t_10 = 1.0f / sqrtf(fmaxf(t_9, ((t_0 * t_0) + (t_3 * t_3))));
float tmp_1;
if (dY_46_u <= 8000000.0f) {
float tmp_2;
if (t_9 >= ((dY_46_v * dY_46_v) * (floorf(h) * floorf(h)))) {
tmp_2 = t_10 * t_2;
} else {
tmp_2 = t_10 * t_3;
}
tmp_1 = tmp_2;
} else if (t_6 >= (t_1 * floorf(w))) {
tmp_1 = t_5 / t_7;
} else {
tmp_1 = t_4 / t_7;
}
return tmp_1;
}
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 * dY_46_u) t_2 = Float32(floor(h) * dX_46_v) t_3 = Float32(floor(h) * dY_46_v) t_4 = Float32(dY_46_v * floor(h)) t_5 = Float32(dX_46_v * floor(h)) t_6 = fma(Float32(t_5 * dX_46_v), floor(h), Float32(Float32(dX_46_u * dX_46_u) * Float32(floor(w) * floor(w)))) t_7 = sqrt(fmax(t_6, fma(t_1, floor(w), Float32(Float32(t_4 * dY_46_v) * floor(h))))) t_8 = Float32(floor(w) * dX_46_u) t_9 = Float32(Float32(t_8 * t_8) + Float32(t_2 * t_2)) t_10 = Float32(Float32(1.0) / sqrt(fmax(t_9, Float32(Float32(t_0 * t_0) + Float32(t_3 * t_3))))) tmp_1 = Float32(0.0) if (dY_46_u <= Float32(8000000.0)) tmp_2 = Float32(0.0) if (t_9 >= Float32(Float32(dY_46_v * dY_46_v) * Float32(floor(h) * floor(h)))) tmp_2 = Float32(t_10 * t_2); else tmp_2 = Float32(t_10 * t_3); end tmp_1 = tmp_2; elseif (t_6 >= Float32(t_1 * floor(w))) tmp_1 = Float32(t_5 / t_7); else tmp_1 = Float32(t_4 / t_7); end return tmp_1 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := t\_0 \cdot dY.u\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_3 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_4 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_5 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_6 := \mathsf{fma}\left(t\_5 \cdot dX.v, \left\lfloor h\right\rfloor , \left(dX.u \cdot dX.u\right) \cdot \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right)\right)\\
t_7 := \sqrt{\mathsf{max}\left(t\_6, \mathsf{fma}\left(t\_1, \left\lfloor w\right\rfloor , \left(t\_4 \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)}\\
t_8 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_9 := t\_8 \cdot t\_8 + t\_2 \cdot t\_2\\
t_10 := \frac{1}{\sqrt{\mathsf{max}\left(t\_9, t\_0 \cdot t\_0 + t\_3 \cdot t\_3\right)}}\\
\mathbf{if}\;dY.u \leq 8000000:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_9 \geq \left(dY.v \cdot dY.v\right) \cdot \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right):\\
\;\;\;\;t\_10 \cdot t\_2\\
\mathbf{else}:\\
\;\;\;\;t\_10 \cdot t\_3\\
\end{array}\\
\mathbf{elif}\;t\_6 \geq t\_1 \cdot \left\lfloor w\right\rfloor :\\
\;\;\;\;\frac{t\_5}{t\_7}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_4}{t\_7}\\
\end{array}
\end{array}
if dY.u < 8e6Initial program 78.0%
Taylor expanded in dY.u around 0
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3269.7
Applied rewrites69.7%
if 8e6 < dY.u Initial program 66.2%
Taylor expanded in dY.u around inf
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3264.6
Applied rewrites64.6%
Applied rewrites64.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* (* dY.v dY.v) t_0))
(t_2 (* (floor w) (floor w)))
(t_3 (fma (* t_2 dY.u) dY.u t_1)))
(if (>= (* (* dX.u dX.u) t_2) t_3)
(*
(/
dX.v
(sqrt
(fmax
(fma
(* (* (floor w) dX.u) dX.u)
(floor w)
(* (* (* dX.v (floor h)) dX.v) (floor h)))
(fma (* dY.u dY.u) t_2 t_1))))
(floor h))
(/
(* (floor h) dY.v)
(sqrt (fmax (fma (* t_2 dX.u) dX.u (* t_0 (* 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(h) * floorf(h);
float t_1 = (dY_46_v * dY_46_v) * t_0;
float t_2 = floorf(w) * floorf(w);
float t_3 = fmaf((t_2 * dY_46_u), dY_46_u, t_1);
float tmp;
if (((dX_46_u * dX_46_u) * t_2) >= t_3) {
tmp = (dX_46_v / sqrtf(fmaxf(fmaf(((floorf(w) * dX_46_u) * dX_46_u), floorf(w), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h))), fmaf((dY_46_u * dY_46_u), t_2, t_1)))) * floorf(h);
} else {
tmp = (floorf(h) * dY_46_v) / sqrtf(fmaxf(fmaf((t_2 * dX_46_u), dX_46_u, (t_0 * (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(h) * floor(h)) t_1 = Float32(Float32(dY_46_v * dY_46_v) * t_0) t_2 = Float32(floor(w) * floor(w)) t_3 = fma(Float32(t_2 * dY_46_u), dY_46_u, t_1) tmp = Float32(0.0) if (Float32(Float32(dX_46_u * dX_46_u) * t_2) >= t_3) tmp = Float32(Float32(dX_46_v / sqrt(fmax(fma(Float32(Float32(floor(w) * dX_46_u) * dX_46_u), floor(w), Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h))), fma(Float32(dY_46_u * dY_46_u), t_2, t_1)))) * floor(h)); else tmp = Float32(Float32(floor(h) * dY_46_v) / sqrt(fmax(fma(Float32(t_2 * dX_46_u), dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))), t_3))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_1 := \left(dY.v \cdot dY.v\right) \cdot t\_0\\
t_2 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_3 := \mathsf{fma}\left(t\_2 \cdot dY.u, dY.u, t\_1\right)\\
\mathbf{if}\;\left(dX.u \cdot dX.u\right) \cdot t\_2 \geq t\_3:\\
\;\;\;\;\frac{dX.v}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot dX.u\right) \cdot dX.u, \left\lfloor w\right\rfloor , \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right), \mathsf{fma}\left(dY.u \cdot dY.u, t\_2, t\_1\right)\right)}} \cdot \left\lfloor h\right\rfloor \\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dY.v}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_2 \cdot dX.u, dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right), t\_3\right)}}\\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.1%
Taylor expanded in dX.u around inf
Applied rewrites64.3%
Applied rewrites64.3%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0
(fma
(* dY.u dY.u)
(* (floor w) (floor w))
(* (* dY.v dY.v) (* (floor h) (floor h)))))
(t_1 (* (* (floor w) dX.u) dX.u))
(t_2
(/
(floor h)
(sqrt
(fmax
(fma t_1 (floor w) (* (* (* dX.v (floor h)) dX.v) (floor h)))
t_0)))))
(if (>= (* t_1 (floor w)) t_0) (* t_2 dX.v) (* t_2 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 = fmaf((dY_46_u * dY_46_u), (floorf(w) * floorf(w)), ((dY_46_v * dY_46_v) * (floorf(h) * floorf(h))));
float t_1 = (floorf(w) * dX_46_u) * dX_46_u;
float t_2 = floorf(h) / sqrtf(fmaxf(fmaf(t_1, floorf(w), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h))), t_0));
float tmp;
if ((t_1 * floorf(w)) >= t_0) {
tmp = t_2 * dX_46_v;
} else {
tmp = t_2 * dY_46_v;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = fma(Float32(dY_46_u * dY_46_u), Float32(floor(w) * floor(w)), Float32(Float32(dY_46_v * dY_46_v) * Float32(floor(h) * floor(h)))) t_1 = Float32(Float32(floor(w) * dX_46_u) * dX_46_u) t_2 = Float32(floor(h) / sqrt(fmax(fma(t_1, floor(w), Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h))), t_0))) tmp = Float32(0.0) if (Float32(t_1 * floor(w)) >= t_0) tmp = Float32(t_2 * dX_46_v); else tmp = Float32(t_2 * dY_46_v); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \mathsf{fma}\left(dY.u \cdot dY.u, \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor , \left(dY.v \cdot dY.v\right) \cdot \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right)\right)\\
t_1 := \left(\left\lfloor w\right\rfloor \cdot dX.u\right) \cdot dX.u\\
t_2 := \frac{\left\lfloor h\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1, \left\lfloor w\right\rfloor , \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right), t\_0\right)}}\\
\mathbf{if}\;t\_1 \cdot \left\lfloor w\right\rfloor \geq t\_0:\\
\;\;\;\;t\_2 \cdot dX.v\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot dY.v\\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.1%
Taylor expanded in dX.u around inf
Applied rewrites64.3%
Applied rewrites64.3%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* (floor w) (floor w)))
(t_2 (* (* dX.u dX.u) t_1))
(t_3 (fma (* t_1 dY.u) dY.u (* (* dY.v dY.v) t_0))))
(if (>= t_2 (* (* dY.u dY.u) t_1))
(/
(* (floor h) dX.v)
(sqrt (fmax (fma (* (* dX.v (floor h)) (floor h)) dX.v t_2) t_3)))
(/
(* (floor h) dY.v)
(sqrt (fmax (fma (* t_1 dX.u) dX.u (* t_0 (* 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(h) * floorf(h);
float t_1 = floorf(w) * floorf(w);
float t_2 = (dX_46_u * dX_46_u) * t_1;
float t_3 = fmaf((t_1 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_0));
float tmp;
if (t_2 >= ((dY_46_u * dY_46_u) * t_1)) {
tmp = (floorf(h) * dX_46_v) / sqrtf(fmaxf(fmaf(((dX_46_v * floorf(h)) * floorf(h)), dX_46_v, t_2), t_3));
} else {
tmp = (floorf(h) * dY_46_v) / sqrtf(fmaxf(fmaf((t_1 * dX_46_u), dX_46_u, (t_0 * (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(h) * floor(h)) t_1 = Float32(floor(w) * floor(w)) t_2 = Float32(Float32(dX_46_u * dX_46_u) * t_1) t_3 = fma(Float32(t_1 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_0)) tmp = Float32(0.0) if (t_2 >= Float32(Float32(dY_46_u * dY_46_u) * t_1)) tmp = Float32(Float32(floor(h) * dX_46_v) / sqrt(fmax(fma(Float32(Float32(dX_46_v * floor(h)) * floor(h)), dX_46_v, t_2), t_3))); else tmp = Float32(Float32(floor(h) * dY_46_v) / sqrt(fmax(fma(Float32(t_1 * dX_46_u), dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))), t_3))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_2 := \left(dX.u \cdot dX.u\right) \cdot t\_1\\
t_3 := \mathsf{fma}\left(t\_1 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_0\right)\\
\mathbf{if}\;t\_2 \geq \left(dY.u \cdot dY.u\right) \cdot t\_1:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dX.v}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot \left\lfloor h\right\rfloor , dX.v, t\_2\right), t\_3\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dY.v}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1 \cdot dX.u, dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right), t\_3\right)}}\\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.1%
Taylor expanded in dX.u around inf
Applied rewrites64.3%
Taylor expanded in dY.u around inf
Applied rewrites58.7%
lift-fma.f32N/A
lift-*.f32N/A
lift-*.f32N/A
Applied rewrites58.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) (floor w)))
(t_1 (* (floor h) (floor h)))
(t_2 (* (* dY.v dY.v) t_1)))
(if (>= (* (* dX.u dX.u) t_0) (* (* dY.u dY.u) t_0))
(*
(floor h)
(/
dX.v
(sqrt
(fmax
(fma (* dX.u dX.u) t_0 (* (* dX.v dX.v) t_1))
(fma (* dY.u dY.u) t_0 t_2)))))
(/
(* (floor h) dY.v)
(sqrt
(fmax
(fma (* t_0 dX.u) dX.u (* t_1 (* dX.v dX.v)))
(fma (* t_0 dY.u) dY.u t_2)))))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(w) * floorf(w);
float t_1 = floorf(h) * floorf(h);
float t_2 = (dY_46_v * dY_46_v) * t_1;
float tmp;
if (((dX_46_u * dX_46_u) * t_0) >= ((dY_46_u * dY_46_u) * t_0)) {
tmp = floorf(h) * (dX_46_v / sqrtf(fmaxf(fmaf((dX_46_u * dX_46_u), t_0, ((dX_46_v * dX_46_v) * t_1)), fmaf((dY_46_u * dY_46_u), t_0, t_2))));
} else {
tmp = (floorf(h) * dY_46_v) / sqrtf(fmaxf(fmaf((t_0 * dX_46_u), dX_46_u, (t_1 * (dX_46_v * dX_46_v))), fmaf((t_0 * dY_46_u), dY_46_u, t_2)));
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * floor(w)) t_1 = Float32(floor(h) * floor(h)) t_2 = Float32(Float32(dY_46_v * dY_46_v) * t_1) tmp = Float32(0.0) if (Float32(Float32(dX_46_u * dX_46_u) * t_0) >= Float32(Float32(dY_46_u * dY_46_u) * t_0)) tmp = Float32(floor(h) * Float32(dX_46_v / sqrt(fmax(fma(Float32(dX_46_u * dX_46_u), t_0, Float32(Float32(dX_46_v * dX_46_v) * t_1)), fma(Float32(dY_46_u * dY_46_u), t_0, t_2))))); else tmp = Float32(Float32(floor(h) * dY_46_v) / sqrt(fmax(fma(Float32(t_0 * dX_46_u), dX_46_u, Float32(t_1 * Float32(dX_46_v * dX_46_v))), fma(Float32(t_0 * dY_46_u), dY_46_u, t_2)))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_1 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_2 := \left(dY.v \cdot dY.v\right) \cdot t\_1\\
\mathbf{if}\;\left(dX.u \cdot dX.u\right) \cdot t\_0 \geq \left(dY.u \cdot dY.u\right) \cdot t\_0:\\
\;\;\;\;\left\lfloor h\right\rfloor \cdot \frac{dX.v}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.u \cdot dX.u, t\_0, \left(dX.v \cdot dX.v\right) \cdot t\_1\right), \mathsf{fma}\left(dY.u \cdot dY.u, t\_0, t\_2\right)\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dY.v}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0 \cdot dX.u, dX.u, t\_1 \cdot \left(dX.v \cdot dX.v\right)\right), \mathsf{fma}\left(t\_0 \cdot dY.u, dY.u, t\_2\right)\right)}}\\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.1%
Taylor expanded in dX.u around inf
Applied rewrites64.3%
Taylor expanded in dY.u around inf
Applied rewrites58.7%
Applied rewrites58.7%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) (floor w)))
(t_1 (* (floor h) (floor h)))
(t_2 (* (* dY.v dY.v) t_1)))
(if (>= (* (* dX.u dX.u) t_0) (* (* dY.u dY.u) t_0))
(*
dX.v
(/
(floor h)
(sqrt
(fmax
(fma (* dX.u dX.u) t_0 (* (* dX.v dX.v) t_1))
(fma (* dY.u dY.u) t_0 t_2)))))
(/
(* (floor h) dY.v)
(sqrt
(fmax
(fma (* t_0 dX.u) dX.u (* t_1 (* dX.v dX.v)))
(fma (* t_0 dY.u) dY.u t_2)))))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(w) * floorf(w);
float t_1 = floorf(h) * floorf(h);
float t_2 = (dY_46_v * dY_46_v) * t_1;
float tmp;
if (((dX_46_u * dX_46_u) * t_0) >= ((dY_46_u * dY_46_u) * t_0)) {
tmp = dX_46_v * (floorf(h) / sqrtf(fmaxf(fmaf((dX_46_u * dX_46_u), t_0, ((dX_46_v * dX_46_v) * t_1)), fmaf((dY_46_u * dY_46_u), t_0, t_2))));
} else {
tmp = (floorf(h) * dY_46_v) / sqrtf(fmaxf(fmaf((t_0 * dX_46_u), dX_46_u, (t_1 * (dX_46_v * dX_46_v))), fmaf((t_0 * dY_46_u), dY_46_u, t_2)));
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * floor(w)) t_1 = Float32(floor(h) * floor(h)) t_2 = Float32(Float32(dY_46_v * dY_46_v) * t_1) tmp = Float32(0.0) if (Float32(Float32(dX_46_u * dX_46_u) * t_0) >= Float32(Float32(dY_46_u * dY_46_u) * t_0)) tmp = Float32(dX_46_v * Float32(floor(h) / sqrt(fmax(fma(Float32(dX_46_u * dX_46_u), t_0, Float32(Float32(dX_46_v * dX_46_v) * t_1)), fma(Float32(dY_46_u * dY_46_u), t_0, t_2))))); else tmp = Float32(Float32(floor(h) * dY_46_v) / sqrt(fmax(fma(Float32(t_0 * dX_46_u), dX_46_u, Float32(t_1 * Float32(dX_46_v * dX_46_v))), fma(Float32(t_0 * dY_46_u), dY_46_u, t_2)))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_1 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_2 := \left(dY.v \cdot dY.v\right) \cdot t\_1\\
\mathbf{if}\;\left(dX.u \cdot dX.u\right) \cdot t\_0 \geq \left(dY.u \cdot dY.u\right) \cdot t\_0:\\
\;\;\;\;dX.v \cdot \frac{\left\lfloor h\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.u \cdot dX.u, t\_0, \left(dX.v \cdot dX.v\right) \cdot t\_1\right), \mathsf{fma}\left(dY.u \cdot dY.u, t\_0, t\_2\right)\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dY.v}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0 \cdot dX.u, dX.u, t\_1 \cdot \left(dX.v \cdot dX.v\right)\right), \mathsf{fma}\left(t\_0 \cdot dY.u, dY.u, t\_2\right)\right)}}\\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.1%
Taylor expanded in dX.u around inf
Applied rewrites64.3%
Taylor expanded in dY.u around inf
Applied rewrites58.7%
Applied rewrites58.7%
herbie shell --seed 2025101
(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))))