
(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 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 (* (floor w) dY.u))
(t_2 (fma t_1 t_1 (* 0.0 dY.v)))
(t_3 (* (floor w) dX.u))
(t_4
(fma
(* (* (floor h) (floor h)) dX.v)
dX.v
(* (* t_3 (floor w)) dX.u)))
(t_5 (* (fma (floor h) (floor h) (/ 0.0 dY.v)) (* dY.v dY.v)))
(t_6 (+ (* t_3 t_3) (* t_0 t_0)))
(t_7 (/ 1.0 (sqrt (fmax t_6 t_5))))
(t_8 (* (floor h) dY.v))
(t_9 (if (>= t_6 t_5) (* t_7 t_0) (* t_7 t_8)))
(t_10 (+ (* t_1 t_1) (* t_8 t_8)))
(t_11 (/ 1.0 (sqrt (fmax t_6 t_10))))
(t_12 (if (>= t_6 t_10) (* t_11 t_0) (* t_11 t_8)))
(t_13 (sqrt (fmax t_4 t_2))))
(if (<= t_12 -0.5)
t_9
(if (<= t_12 0.004999999888241291)
(if (>= t_4 t_2) (/ (* 1.0 t_0) t_13) (/ (* 1.0 t_8) t_13))
t_9))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(w) * dY_46_u;
float t_2 = fmaf(t_1, t_1, (0.0f * dY_46_v));
float t_3 = floorf(w) * dX_46_u;
float t_4 = fmaf(((floorf(h) * floorf(h)) * dX_46_v), dX_46_v, ((t_3 * floorf(w)) * dX_46_u));
float t_5 = fmaf(floorf(h), floorf(h), (0.0f / dY_46_v)) * (dY_46_v * dY_46_v);
float t_6 = (t_3 * t_3) + (t_0 * t_0);
float t_7 = 1.0f / sqrtf(fmaxf(t_6, t_5));
float t_8 = floorf(h) * dY_46_v;
float tmp;
if (t_6 >= t_5) {
tmp = t_7 * t_0;
} else {
tmp = t_7 * t_8;
}
float t_9 = tmp;
float t_10 = (t_1 * t_1) + (t_8 * t_8);
float t_11 = 1.0f / sqrtf(fmaxf(t_6, t_10));
float tmp_1;
if (t_6 >= t_10) {
tmp_1 = t_11 * t_0;
} else {
tmp_1 = t_11 * t_8;
}
float t_12 = tmp_1;
float t_13 = sqrtf(fmaxf(t_4, t_2));
float tmp_2;
if (t_12 <= -0.5f) {
tmp_2 = t_9;
} else if (t_12 <= 0.004999999888241291f) {
float tmp_3;
if (t_4 >= t_2) {
tmp_3 = (1.0f * t_0) / t_13;
} else {
tmp_3 = (1.0f * t_8) / t_13;
}
tmp_2 = tmp_3;
} else {
tmp_2 = t_9;
}
return tmp_2;
}
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 = fma(t_1, t_1, Float32(Float32(0.0) * dY_46_v)) t_3 = Float32(floor(w) * dX_46_u) t_4 = fma(Float32(Float32(floor(h) * floor(h)) * dX_46_v), dX_46_v, Float32(Float32(t_3 * floor(w)) * dX_46_u)) t_5 = Float32(fma(floor(h), floor(h), Float32(Float32(0.0) / dY_46_v)) * Float32(dY_46_v * dY_46_v)) t_6 = Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)) t_7 = Float32(Float32(1.0) / sqrt(fmax(t_6, t_5))) t_8 = Float32(floor(h) * dY_46_v) tmp = Float32(0.0) if (t_6 >= t_5) tmp = Float32(t_7 * t_0); else tmp = Float32(t_7 * t_8); end t_9 = tmp t_10 = Float32(Float32(t_1 * t_1) + Float32(t_8 * t_8)) t_11 = Float32(Float32(1.0) / sqrt(fmax(t_6, t_10))) tmp_1 = Float32(0.0) if (t_6 >= t_10) tmp_1 = Float32(t_11 * t_0); else tmp_1 = Float32(t_11 * t_8); end t_12 = tmp_1 t_13 = sqrt(fmax(t_4, t_2)) tmp_2 = Float32(0.0) if (t_12 <= Float32(-0.5)) tmp_2 = t_9; elseif (t_12 <= Float32(0.004999999888241291)) tmp_3 = Float32(0.0) if (t_4 >= t_2) tmp_3 = Float32(Float32(Float32(1.0) * t_0) / t_13); else tmp_3 = Float32(Float32(Float32(1.0) * t_8) / t_13); end tmp_2 = tmp_3; else tmp_2 = t_9; end return tmp_2 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 := \mathsf{fma}\left(t\_1, t\_1, 0 \cdot dY.v\right)\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := \mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, dX.v, \left(t\_3 \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right)\\
t_5 := \mathsf{fma}\left(\left\lfloor h\right\rfloor , \left\lfloor h\right\rfloor , \frac{0}{dY.v}\right) \cdot \left(dY.v \cdot dY.v\right)\\
t_6 := t\_3 \cdot t\_3 + t\_0 \cdot t\_0\\
t_7 := \frac{1}{\sqrt{\mathsf{max}\left(t\_6, t\_5\right)}}\\
t_8 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_9 := \begin{array}{l}
\mathbf{if}\;t\_6 \geq t\_5:\\
\;\;\;\;t\_7 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_7 \cdot t\_8\\
\end{array}\\
t_10 := t\_1 \cdot t\_1 + t\_8 \cdot t\_8\\
t_11 := \frac{1}{\sqrt{\mathsf{max}\left(t\_6, t\_10\right)}}\\
t_12 := \begin{array}{l}
\mathbf{if}\;t\_6 \geq t\_10:\\
\;\;\;\;t\_11 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_11 \cdot t\_8\\
\end{array}\\
t_13 := \sqrt{\mathsf{max}\left(t\_4, t\_2\right)}\\
\mathbf{if}\;t\_12 \leq -0.5:\\
\;\;\;\;t\_9\\
\mathbf{elif}\;t\_12 \leq 0.004999999888241291:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_4 \geq t\_2:\\
\;\;\;\;\frac{1 \cdot t\_0}{t\_13}\\
\mathbf{else}:\\
\;\;\;\;\frac{1 \cdot t\_8}{t\_13}\\
\end{array}\\
\mathbf{else}:\\
\;\;\;\;t\_9\\
\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))) < -0.5 or 0.00499999989 < (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 76.2%
Applied rewrites69.1%
Applied rewrites69.1%
Applied rewrites45.9%
Taylor expanded in dY.v around 0
+-commutativeN/A
*-commutativeN/A
pow2N/A
unpow2N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites43.9%
Taylor expanded in dY.v around 0
+-commutativeN/A
*-commutativeN/A
pow2N/A
unpow2N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites52.9%
Taylor expanded in dY.v around 0
+-commutativeN/A
*-commutativeN/A
pow2N/A
unpow2N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites62.3%
Taylor expanded in dY.v around inf
Applied rewrites49.8%
Taylor expanded in dY.v around inf
Applied rewrites49.4%
Taylor expanded in dY.v around inf
Applied rewrites59.2%
if -0.5 < (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))) < 0.00499999989Initial program 76.2%
Applied rewrites69.1%
Applied rewrites69.1%
Applied rewrites45.9%
Taylor expanded in dY.v around 0
+-commutativeN/A
*-commutativeN/A
pow2N/A
unpow2N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites43.9%
Taylor expanded in dY.v around 0
+-commutativeN/A
*-commutativeN/A
pow2N/A
unpow2N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites52.9%
Taylor expanded in dY.v around 0
+-commutativeN/A
*-commutativeN/A
pow2N/A
unpow2N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites62.3%
Applied rewrites62.4%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dY.v))
(t_1 (* (floor w) (floor w)))
(t_2 (fma t_0 t_0 (* (* dY.u dY.u) t_1)))
(t_3 (fma (* t_1 dX.u) dX.u (* (* (floor h) (floor h)) (* dX.v dX.v))))
(t_4 (sqrt (fmax t_3 t_2))))
(if (>= t_3 t_2) (/ (* (floor h) dX.v) t_4) (/ t_0 t_4))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dY_46_v;
float t_1 = floorf(w) * floorf(w);
float t_2 = fmaf(t_0, t_0, ((dY_46_u * dY_46_u) * t_1));
float t_3 = fmaf((t_1 * dX_46_u), dX_46_u, ((floorf(h) * floorf(h)) * (dX_46_v * dX_46_v)));
float t_4 = sqrtf(fmaxf(t_3, t_2));
float tmp;
if (t_3 >= t_2) {
tmp = (floorf(h) * dX_46_v) / t_4;
} else {
tmp = t_0 / 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) * dY_46_v) t_1 = Float32(floor(w) * floor(w)) t_2 = fma(t_0, t_0, Float32(Float32(dY_46_u * dY_46_u) * t_1)) t_3 = fma(Float32(t_1 * dX_46_u), dX_46_u, Float32(Float32(floor(h) * floor(h)) * Float32(dX_46_v * dX_46_v))) t_4 = sqrt(fmax(t_3, t_2)) tmp = Float32(0.0) if (t_3 >= t_2) tmp = Float32(Float32(floor(h) * dX_46_v) / t_4); else tmp = Float32(t_0 / t_4); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_1 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_2 := \mathsf{fma}\left(t\_0, t\_0, \left(dY.u \cdot dY.u\right) \cdot t\_1\right)\\
t_3 := \mathsf{fma}\left(t\_1 \cdot dX.u, dX.u, \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot \left(dX.v \cdot dX.v\right)\right)\\
t_4 := \sqrt{\mathsf{max}\left(t\_3, t\_2\right)}\\
\mathbf{if}\;t\_3 \geq t\_2:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dX.v}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_0}{t\_4}\\
\end{array}
\end{array}
Initial program 76.2%
Applied rewrites76.3%
lift-fma.f32N/A
+-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
Applied rewrites76.3%
lift-fma.f32N/A
+-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
Applied rewrites76.3%
lift-fma.f32N/A
+-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
Applied rewrites76.4%
(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) (floor w)))
(t_2 (fma (* t_1 dY.u) dY.u (* (* dY.v dY.v) (* (floor h) (floor h)))))
(t_3 (fma (* t_1 dX.u) dX.u (* t_0 t_0)))
(t_4 (sqrt (fmax t_3 t_2))))
(if (>= t_3 t_2) (/ t_0 t_4) (/ (* (floor h) dY.v) 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) * floorf(w);
float t_2 = fmaf((t_1 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * (floorf(h) * floorf(h))));
float t_3 = fmaf((t_1 * dX_46_u), dX_46_u, (t_0 * t_0));
float t_4 = sqrtf(fmaxf(t_3, t_2));
float tmp;
if (t_3 >= t_2) {
tmp = t_0 / t_4;
} else {
tmp = (floorf(h) * dY_46_v) / 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) * floor(w)) t_2 = fma(Float32(t_1 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * Float32(floor(h) * floor(h)))) t_3 = fma(Float32(t_1 * dX_46_u), dX_46_u, Float32(t_0 * t_0)) t_4 = sqrt(fmax(t_3, t_2)) tmp = Float32(0.0) if (t_3 >= t_2) tmp = Float32(t_0 / t_4); else tmp = Float32(Float32(floor(h) * dY_46_v) / 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\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 \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right)\right)\\
t_3 := \mathsf{fma}\left(t\_1 \cdot dX.u, dX.u, t\_0 \cdot t\_0\right)\\
t_4 := \sqrt{\mathsf{max}\left(t\_3, t\_2\right)}\\
\mathbf{if}\;t\_3 \geq t\_2:\\
\;\;\;\;\frac{t\_0}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dY.v}{t\_4}\\
\end{array}
\end{array}
Initial program 76.2%
Applied rewrites76.3%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3276.3
Applied rewrites76.3%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3276.4
Applied rewrites76.4%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3276.4
Applied rewrites76.4%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (fma (* (* (floor w) (floor w)) dY.u) dY.u (* (* dY.v dY.v) t_0)))
(t_2 (* (floor w) dX.u))
(t_3 (fma t_2 t_2 (* (* t_0 dX.v) dX.v)))
(t_4 (sqrt (fmax t_3 t_1))))
(if (>= t_3 t_1) (/ (* (floor h) dX.v) t_4) (/ (* (floor h) dY.v) 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) * floorf(h);
float t_1 = fmaf(((floorf(w) * floorf(w)) * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_0));
float t_2 = floorf(w) * dX_46_u;
float t_3 = fmaf(t_2, t_2, ((t_0 * dX_46_v) * dX_46_v));
float t_4 = sqrtf(fmaxf(t_3, t_1));
float tmp;
if (t_3 >= t_1) {
tmp = (floorf(h) * dX_46_v) / t_4;
} else {
tmp = (floorf(h) * dY_46_v) / 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) * floor(h)) t_1 = fma(Float32(Float32(floor(w) * floor(w)) * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_0)) t_2 = Float32(floor(w) * dX_46_u) t_3 = fma(t_2, t_2, Float32(Float32(t_0 * dX_46_v) * dX_46_v)) t_4 = sqrt(fmax(t_3, t_1)) tmp = Float32(0.0) if (t_3 >= t_1) tmp = Float32(Float32(floor(h) * dX_46_v) / t_4); else tmp = Float32(Float32(floor(h) * dY_46_v) / t_4); 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 := \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_0\right)\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := \mathsf{fma}\left(t\_2, t\_2, \left(t\_0 \cdot dX.v\right) \cdot dX.v\right)\\
t_4 := \sqrt{\mathsf{max}\left(t\_3, t\_1\right)}\\
\mathbf{if}\;t\_3 \geq t\_1:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dX.v}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dY.v}{t\_4}\\
\end{array}
\end{array}
Initial program 76.2%
Applied rewrites76.3%
lift-fma.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
Applied rewrites76.3%
lift-fma.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
Applied rewrites76.3%
lift-fma.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
Applied rewrites76.3%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (fma t_0 t_0 (* 0.0 dY.v)))
(t_2
(fma
(* (* (floor h) (floor h)) dX.v)
dX.v
(* (* (* (floor w) dX.u) (floor w)) dX.u)))
(t_3 (sqrt (fmax t_2 t_1))))
(if (>= t_2 t_1)
(/ (* 1.0 (* (floor h) dX.v)) t_3)
(/ (* 1.0 (* (floor h) dY.v)) t_3))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(w) * dY_46_u;
float t_1 = fmaf(t_0, t_0, (0.0f * dY_46_v));
float t_2 = fmaf(((floorf(h) * floorf(h)) * dX_46_v), dX_46_v, (((floorf(w) * dX_46_u) * floorf(w)) * dX_46_u));
float t_3 = sqrtf(fmaxf(t_2, t_1));
float tmp;
if (t_2 >= t_1) {
tmp = (1.0f * (floorf(h) * dX_46_v)) / t_3;
} else {
tmp = (1.0f * (floorf(h) * dY_46_v)) / t_3;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dY_46_u) t_1 = fma(t_0, t_0, Float32(Float32(0.0) * dY_46_v)) t_2 = fma(Float32(Float32(floor(h) * floor(h)) * dX_46_v), dX_46_v, Float32(Float32(Float32(floor(w) * dX_46_u) * floor(w)) * dX_46_u)) t_3 = sqrt(fmax(t_2, t_1)) tmp = Float32(0.0) if (t_2 >= t_1) tmp = Float32(Float32(Float32(1.0) * Float32(floor(h) * dX_46_v)) / t_3); else tmp = Float32(Float32(Float32(1.0) * Float32(floor(h) * dY_46_v)) / t_3); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \mathsf{fma}\left(t\_0, t\_0, 0 \cdot dY.v\right)\\
t_2 := \mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, dX.v, \left(\left(\left\lfloor w\right\rfloor \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right)\\
t_3 := \sqrt{\mathsf{max}\left(t\_2, t\_1\right)}\\
\mathbf{if}\;t\_2 \geq t\_1:\\
\;\;\;\;\frac{1 \cdot \left(\left\lfloor h\right\rfloor \cdot dX.v\right)}{t\_3}\\
\mathbf{else}:\\
\;\;\;\;\frac{1 \cdot \left(\left\lfloor h\right\rfloor \cdot dY.v\right)}{t\_3}\\
\end{array}
\end{array}
Initial program 76.2%
Applied rewrites69.1%
Applied rewrites69.1%
Applied rewrites45.9%
Taylor expanded in dY.v around 0
+-commutativeN/A
*-commutativeN/A
pow2N/A
unpow2N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites43.9%
Taylor expanded in dY.v around 0
+-commutativeN/A
*-commutativeN/A
pow2N/A
unpow2N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites52.9%
Taylor expanded in dY.v around 0
+-commutativeN/A
*-commutativeN/A
pow2N/A
unpow2N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites62.3%
Applied rewrites62.4%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (fma (* (* (floor w) dY.u) (floor w)) dY.u 0.0))
(t_2 (* (floor w) dX.u))
(t_3 (+ (* t_2 t_2) (* t_0 t_0)))
(t_4 (/ 1.0 (sqrt (fmax t_3 t_1)))))
(if (>= t_3 t_1) (* t_4 t_0) (* t_4 (* (floor h) dY.v)))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = fmaf(((floorf(w) * dY_46_u) * floorf(w)), dY_46_u, 0.0f);
float t_2 = floorf(w) * dX_46_u;
float t_3 = (t_2 * t_2) + (t_0 * t_0);
float t_4 = 1.0f / sqrtf(fmaxf(t_3, t_1));
float tmp;
if (t_3 >= t_1) {
tmp = t_4 * t_0;
} else {
tmp = t_4 * (floorf(h) * dY_46_v);
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = fma(Float32(Float32(floor(w) * dY_46_u) * floor(w)), dY_46_u, Float32(0.0)) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) t_4 = Float32(Float32(1.0) / sqrt(fmax(t_3, t_1))) tmp = Float32(0.0) if (t_3 >= t_1) tmp = Float32(t_4 * t_0); else tmp = Float32(t_4 * Float32(floor(h) * dY_46_v)); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor , dY.u, 0\right)\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := t\_2 \cdot t\_2 + t\_0 \cdot t\_0\\
t_4 := \frac{1}{\sqrt{\mathsf{max}\left(t\_3, t\_1\right)}}\\
\mathbf{if}\;t\_3 \geq t\_1:\\
\;\;\;\;t\_4 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_4 \cdot \left(\left\lfloor h\right\rfloor \cdot dY.v\right)\\
\end{array}
\end{array}
Initial program 76.2%
Applied rewrites69.1%
Applied rewrites69.1%
Applied rewrites45.9%
Taylor expanded in dY.v around 0
+-commutativeN/A
*-commutativeN/A
pow2N/A
unpow2N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites43.9%
Taylor expanded in dY.v around 0
+-commutativeN/A
*-commutativeN/A
pow2N/A
unpow2N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites52.9%
Taylor expanded in dY.v around 0
+-commutativeN/A
*-commutativeN/A
pow2N/A
unpow2N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites62.3%
Taylor expanded in w around 0
Applied rewrites62.3%
Taylor expanded in w around 0
Applied rewrites62.3%
Taylor expanded in w around 0
Applied rewrites62.3%
herbie shell --seed 2025134
(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))))