int WolfeLineSearch(FunctorType &func,
Scalar &alpha,
XType &x1, Scalar &f1, XType &gradx1,
const XType &p,
const XType &x0, const Scalar &f0, const XType &gradx0,
const Scalar &c1, const Scalar &c2,
const Scalar &minAlpha, const Scalar &maxLSIts,
const Scalar &maxLSRestarts) {
const Scalar dfp(gradx0.dot(p));
const Scalar c1dfp(c1*dfp);
const Scalar c2dfp(c2*dfp);
Scalar alpha0(minAlpha);
Scalar alpha1(alpha);
Scalar prevF(f0);
XType prevDF(gradx0);
Scalar prevDFp(dfp);
Scalar newDFp;
int retCode = 0, nits = 0, lsRestarts = 0, ret;
while (1) {
if (nits >= maxLSIts) {
retCode = 1;
break;
}
x1.noalias() = x0 + alpha1 * p;
ret = func(x1, f1, gradx1);
if (ret != 0) {
if (lsRestarts >= maxLSRestarts) {
retCode = 1;
break;
}
alpha1 = 0.5 * (alpha0 + alpha1);
lsRestarts++;
continue;
}
lsRestarts = 0;
newDFp = gradx1.dot(p);
if ((f1 > f0 + alpha * c1dfp) || (f1 >= prevF && nits > 0)) {
retCode = WolfLSZoom(alpha, x1, f1, gradx1,
func,
x0, f0, dfp,
c1dfp, c2dfp, p,
alpha0, prevF, prevDFp,
alpha1, f1, newDFp,
1e-16);
break;
}
if (std::fabs(newDFp) <= -c2dfp) {
alpha = alpha1;
break;
}
if (newDFp >= 0) {
retCode = WolfLSZoom(alpha, x1, f1, gradx1,
func,
x0, f0, dfp,
c1dfp, c2dfp, p,
alpha1, f1, newDFp,
alpha0, prevF, prevDFp,
1e-16);
break;
}
alpha0 = alpha1;
prevF = f1;
std::swap(prevDF, gradx1);
prevDFp = newDFp;
alpha1 *= 10.0;
nits++;
}
return retCode;
}
void GSRendererHW::RoundSpriteOffset()
{
//#define DEBUG_U
//#define DEBUG_V
#if defined(DEBUG_V) || defined(DEBUG_U)
bool debug = linear;
#endif
size_t count = m_vertex.next;
GSVertex* v = &m_vertex.buff[0];
for(size_t i = 0; i < count; i += 2) {
// Performance note: if it had any impact on perf, someone would port it to SSE (AKA GSVector)
// Compute the coordinate of first and last texels (in native with a linear filtering)
int ox = m_context->XYOFFSET.OFX;
int X0 = v[i].XYZ.X - ox;
int X1 = v[i+1].XYZ.X - ox;
int Lx = (v[i+1].XYZ.X - v[i].XYZ.X);
float ax0 = alpha0(Lx, X0, X1);
float ax1 = alpha1(Lx, X0, X1);
int tx0 = Interpolate_UV(ax0, v[i].U, v[i+1].U);
int tx1 = Interpolate_UV(ax1, v[i].U, v[i+1].U);
#ifdef DEBUG_U
if (debug) {
fprintf(stderr, "u0:%d and u1:%d\n", v[i].U, v[i+1].U);
fprintf(stderr, "a0:%f and a1:%f\n", ax0, ax1);
fprintf(stderr, "t0:%d and t1:%d\n", tx0, tx1);
}
#endif
int oy = m_context->XYOFFSET.OFY;
int Y0 = v[i].XYZ.Y - oy;
int Y1 = v[i+1].XYZ.Y - oy;
int Ly = (v[i+1].XYZ.Y - v[i].XYZ.Y);
float ay0 = alpha0(Ly, Y0, Y1);
float ay1 = alpha1(Ly, Y0, Y1);
int ty0 = Interpolate_UV(ay0, v[i].V, v[i+1].V);
int ty1 = Interpolate_UV(ay1, v[i].V, v[i+1].V);
#ifdef DEBUG_V
if (debug) {
fprintf(stderr, "v0:%d and v1:%d\n", v[i].V, v[i+1].V);
fprintf(stderr, "a0:%f and a1:%f\n", ay0, ay1);
fprintf(stderr, "t0:%d and t1:%d\n", ty0, ty1);
}
#endif
#ifdef DEBUG_U
if (debug)
fprintf(stderr, "GREP_BEFORE %d => %d\n", v[i].U, v[i+1].U);
#endif
#ifdef DEBUG_V
if (debug)
fprintf(stderr, "GREP_BEFORE %d => %d\n", v[i].V, v[i+1].V);
#endif
#if 1
// Use rounded value of the newly computed texture coordinate. It ensures
// that sampling will remains inside texture boundary
//
// Note for bilinear: by definition it will never work correctly! A sligh modification
// of interpolation migth trigger a discard (with alpha testing)
// Let's use something simple that correct really bad case (for a couple of 2D games).
// I hope it won't create too much glitches.
if (linear) {
int Lu = v[i+1].U - v[i].U;
// Note 32 is based on taisho-mononoke
if ((Lu > 0) && (Lu <= (Lx+32))) {
v[i+1].U -= 8;
}
} else {
if (tx0 <= tx1) {
v[i].U = tx0;
v[i+1].U = tx1 + 16;
} else {
v[i].U = tx0 + 15;
v[i+1].U = tx1;
}
}
#endif
#if 1
if (linear) {
int Lv = v[i+1].V - v[i].V;
if ((Lv > 0) && (Lv <= (Ly+32))) {
v[i+1].V -= 8;
}
} else {
if (ty0 <= ty1) {
v[i].V = ty0;
v[i+1].V = ty1 + 16;
} else {
v[i].V = ty0 + 15;
v[i+1].V = ty1;
}
}
#endif
#ifdef DEBUG_U
if (debug)
fprintf(stderr, "GREP_AFTER %d => %d\n\n", v[i].U, v[i+1].U);
#endif
#ifdef DEBUG_V
if (debug)
fprintf(stderr, "GREP_AFTER %d => %d\n\n", v[i].V, v[i+1].V);
#endif
}
}