void BLI_ewa_filter(const int width, const int height,
const bool intpol,
const bool use_alpha,
const float uv[2],
const float du[2],
const float dv[2],
ewa_filter_read_pixel_cb read_pixel_cb,
void *userdata,
float result[4])
{
/* scaling dxt/dyt by full resolution can cause overflow because of huge A/B/C and esp. F values,
* scaling by aspect ratio alone does the opposite, so try something in between instead... */
const float ff2 = (float)width, ff = sqrtf(ff2), q = (float)height / ff;
const float Ux = du[0] * ff, Vx = du[1] * q, Uy = dv[0] * ff, Vy = dv[1] * q;
float A = Vx * Vx + Vy * Vy;
float B = -2.0f * (Ux * Vx + Uy * Vy);
float C = Ux * Ux + Uy * Uy;
float F = A * C - B * B * 0.25f;
float a, b, th, ecc, a2, b2, ue, ve, U0, V0, DDQ, U, ac1, ac2, BU, d;
int u, v, u1, u2, v1, v2;
/* The so-called 'high' quality ewa method simply adds a constant of 1 to both A & C,
* so the ellipse always covers at least some texels. But since the filter is now always larger,
* it also means that everywhere else it's also more blurry then ideally should be the case.
* So instead here the ellipse radii are modified instead whenever either is too low.
* Use a different radius based on interpolation switch, just enough to anti-alias when interpolation is off,
* and slightly larger to make result a bit smoother than bilinear interpolation when interpolation is on
* (minimum values: const float rmin = intpol ? 1.f : 0.5f;) */
const float rmin = (intpol ? 1.5625f : 0.765625f) / ff2;
BLI_ewa_imp2radangle(A, B, C, F, &a, &b, &th, &ecc);
if ((b2 = b * b) < rmin) {
if ((a2 = a * a) < rmin) {
B = 0.0f;
A = C = rmin;
F = A * C;
}
else {
b2 = rmin;
radangle2imp(a2, b2, th, &A, &B, &C, &F);
}
}
ue = ff * sqrtf(C);
ve = ff * sqrtf(A);
d = (float)(EWA_MAXIDX + 1) / (F * ff2);
A *= d;
B *= d;
C *= d;
U0 = uv[0] * (float)width;
V0 = uv[1] * (float)height;
u1 = (int)(floorf(U0 - ue));
u2 = (int)(ceilf(U0 + ue));
v1 = (int)(floorf(V0 - ve));
v2 = (int)(ceilf(V0 + ve));
/* sane clamping to avoid unnecessarily huge loops */
/* note: if eccentricity gets clamped (see above),
* the ue/ve limits can also be lowered accordingly
*/
if (U0 - (float)u1 > EWA_MAXIDX) u1 = (int)U0 - EWA_MAXIDX;
if ((float)u2 - U0 > EWA_MAXIDX) u2 = (int)U0 + EWA_MAXIDX;
if (V0 - (float)v1 > EWA_MAXIDX) v1 = (int)V0 - EWA_MAXIDX;
if ((float)v2 - V0 > EWA_MAXIDX) v2 = (int)V0 + EWA_MAXIDX;
/* Early output check for cases the whole region is outside of the buffer. */
if ((u2 < 0 || u1 >= width) || (v2 < 0 || v1 >= height)) {
zero_v4(result);
return;
}
U0 -= 0.5f;
V0 -= 0.5f;
DDQ = 2.0f * A;
U = (float)u1 - U0;
ac1 = A * (2.0f * U + 1.0f);
ac2 = A * U * U;
BU = B * U;
d = 0.0f;
zero_v4(result);
for (v = v1; v <= v2; ++v) {
const float V = (float)v - V0;
float DQ = ac1 + B * V;
float Q = (C * V + BU) * V + ac2;
for (u = u1; u <= u2; ++u) {
if (Q < (float)(EWA_MAXIDX + 1)) {
float tc[4];
const float wt = EWA_WTS[(Q < 0.0f) ? 0 : (unsigned int)Q];
read_pixel_cb(userdata, u, v, tc);
madd_v3_v3fl(result, tc, wt);
result[3] += use_alpha ? tc[3] * wt : 0.0f;
d += wt;
}
Q += DQ;
DQ += DDQ;
}
}
/* d should hopefully never be zero anymore */
d = 1.0f / d;
mul_v3_fl(result, d);
/* clipping can be ignored if alpha used, texr->ta already includes filtered edge */
result[3] = use_alpha ? result[3] * d : 1.0f;
}
static int imagewraposa_aniso(Tex *tex, Image *ima, ImBuf *ibuf, const float texvec[3], float dxt[2], float dyt[2], TexResult *texres, struct ImagePool *pool, const bool skip_load_image)
{
TexResult texr;
float fx, fy, minx, maxx, miny, maxy;
float maxd, val1, val2, val3;
int curmap, retval, intpol, extflag = 0;
afdata_t AFD;
void (*filterfunc)(TexResult*, ImBuf*, float, float, afdata_t*);
switch (tex->texfilter) {
case TXF_EWA:
filterfunc = ewa_eval;
break;
case TXF_FELINE:
filterfunc = feline_eval;
break;
case TXF_AREA:
default:
filterfunc = area_sample;
}
texres->tin = texres->ta = texres->tr = texres->tg = texres->tb = 0.f;
/* we need to set retval OK, otherwise texture code generates normals itself... */
retval = texres->nor ? 3 : 1;
/* quick tests */
if (ibuf==NULL && ima==NULL) return retval;
if (ima) { /* hack for icon render */
if (skip_load_image && !BKE_image_has_loaded_ibuf(ima)) {
return retval;
}
ibuf = BKE_image_pool_acquire_ibuf(ima, &tex->iuser, pool);
}
if ((ibuf == NULL) || ((ibuf->rect == NULL) && (ibuf->rect_float == NULL))) {
if (ima)
BKE_image_pool_release_ibuf(ima, ibuf, pool);
return retval;
}
if (ima) {
ima->flag |= IMA_USED_FOR_RENDER;
}
/* mipmap test */
image_mipmap_test(tex, ibuf);
if (ima) {
if ((tex->imaflag & TEX_USEALPHA) && (ima->flag & IMA_IGNORE_ALPHA) == 0) {
if ((tex->imaflag & TEX_CALCALPHA) == 0) {
texres->talpha = 1;
}
}
}
texr.talpha = texres->talpha;
if (tex->imaflag & TEX_IMAROT) {
fy = texvec[0];
fx = texvec[1];
}
else {
fx = texvec[0];
fy = texvec[1];
}
if (ibuf->flags & IB_fields) {
if (R.r.mode & R_FIELDS) { /* field render */
if (R.flag & R_SEC_FIELD) { /* correction for 2nd field */
/* fac1= 0.5/( (float)ibuf->y ); */
/* fy-= fac1; */
}
else /* first field */
fy += 0.5f/( (float)ibuf->y );
}
}
/* pixel coordinates */
minx = min_fff(dxt[0], dyt[0], dxt[0] + dyt[0]);
maxx = max_fff(dxt[0], dyt[0], dxt[0] + dyt[0]);
miny = min_fff(dxt[1], dyt[1], dxt[1] + dyt[1]);
maxy = max_fff(dxt[1], dyt[1], dxt[1] + dyt[1]);
/* tex_sharper has been removed */
minx = (maxx - minx)*0.5f;
miny = (maxy - miny)*0.5f;
if (tex->imaflag & TEX_FILTER_MIN) {
/* make sure the filtersize is minimal in pixels (normal, ref map can have miniature pixel dx/dy) */
const float addval = (0.5f * tex->filtersize) / (float)MIN2(ibuf->x, ibuf->y);
if (addval > minx) minx = addval;
if (addval > miny) miny = addval;
}
else if (tex->filtersize != 1.f) {
minx *= tex->filtersize;
miny *= tex->filtersize;
dxt[0] *= tex->filtersize;
dxt[1] *= tex->filtersize;
dyt[0] *= tex->filtersize;
dyt[1] *= tex->filtersize;
}
if (tex->imaflag & TEX_IMAROT) {
float t;
SWAP(float, minx, miny);
/* must rotate dxt/dyt 90 deg
* yet another blender problem is that swapping X/Y axes (or any tex proj switches) should do something similar,
* but it doesn't, it only swaps coords, so filter area will be incorrect in those cases. */
t = dxt[0];
dxt[0] = dxt[1];
dxt[1] = -t;
t = dyt[0];
dyt[0] = dyt[1];
dyt[1] = -t;
}
/* side faces of unit-cube */
minx = (minx > 0.25f) ? 0.25f : ((minx < 1e-5f) ? 1e-5f : minx);
miny = (miny > 0.25f) ? 0.25f : ((miny < 1e-5f) ? 1e-5f : miny);
/* repeat and clip */
if (tex->extend == TEX_REPEAT) {
if ((tex->flag & (TEX_REPEAT_XMIR | TEX_REPEAT_YMIR)) == (TEX_REPEAT_XMIR | TEX_REPEAT_YMIR))
extflag = TXC_EXTD;
else if (tex->flag & TEX_REPEAT_XMIR)
extflag = TXC_XMIR;
else if (tex->flag & TEX_REPEAT_YMIR)
extflag = TXC_YMIR;
else
extflag = TXC_REPT;
}
else if (tex->extend == TEX_EXTEND)
extflag = TXC_EXTD;
if (tex->extend == TEX_CHECKER) {
int xs = (int)floorf(fx), ys = (int)floorf(fy);
/* both checkers available, no boundary exceptions, checkerdist will eat aliasing */
if ((tex->flag & TEX_CHECKER_ODD) && (tex->flag & TEX_CHECKER_EVEN)) {
fx -= xs;
fy -= ys;
}
else if ((tex->flag & TEX_CHECKER_ODD) == 0 &&
(tex->flag & TEX_CHECKER_EVEN) == 0)
{
if (ima)
BKE_image_pool_release_ibuf(ima, ibuf, pool);
return retval;
}
else {
int xs1 = (int)floorf(fx - minx);
int ys1 = (int)floorf(fy - miny);
int xs2 = (int)floorf(fx + minx);
int ys2 = (int)floorf(fy + miny);
if ((xs1 != xs2) || (ys1 != ys2)) {
if (tex->flag & TEX_CHECKER_ODD) {
fx -= ((xs1 + ys) & 1) ? xs2 : xs1;
fy -= ((ys1 + xs) & 1) ? ys2 : ys1;
}
if (tex->flag & TEX_CHECKER_EVEN) {
fx -= ((xs1 + ys) & 1) ? xs1 : xs2;
fy -= ((ys1 + xs) & 1) ? ys1 : ys2;
}
}
else {
if ((tex->flag & TEX_CHECKER_ODD) == 0 && ((xs + ys) & 1) == 0) {
if (ima)
BKE_image_pool_release_ibuf(ima, ibuf, pool);
return retval;
}
if ((tex->flag & TEX_CHECKER_EVEN) == 0 && (xs + ys) & 1) {
if (ima)
BKE_image_pool_release_ibuf(ima, ibuf, pool);
return retval;
}
fx -= xs;
fy -= ys;
}
}
/* scale around center, (0.5, 0.5) */
if (tex->checkerdist < 1.f) {
const float omcd = 1.f / (1.f - tex->checkerdist);
fx = (fx - 0.5f)*omcd + 0.5f;
fy = (fy - 0.5f)*omcd + 0.5f;
minx *= omcd;
miny *= omcd;
}
}
if (tex->extend == TEX_CLIPCUBE) {
if ((fx + minx) < 0.f || (fy + miny) < 0.f || (fx - minx) > 1.f || (fy - miny) > 1.f || texvec[2] < -1.f || texvec[2] > 1.f) {
if (ima)
BKE_image_pool_release_ibuf(ima, ibuf, pool);
return retval;
}
}
else if (tex->extend == TEX_CLIP || tex->extend == TEX_CHECKER) {
if ((fx + minx) < 0.f || (fy + miny) < 0.f || (fx - minx) > 1.f || (fy - miny) > 1.f) {
if (ima)
BKE_image_pool_release_ibuf(ima, ibuf, pool);
return retval;
}
}
else {
if (tex->extend == TEX_EXTEND) {
fx = (fx > 1.f) ? 1.f : ((fx < 0.f) ? 0.f : fx);
fy = (fy > 1.f) ? 1.f : ((fy < 0.f) ? 0.f : fy);
}
else {
fx -= floorf(fx);
fy -= floorf(fy);
}
}
intpol = tex->imaflag & TEX_INTERPOL;
/* warning no return! */
if ((R.flag & R_SEC_FIELD) && (ibuf->flags & IB_fields))
ibuf->rect += ibuf->x*ibuf->y;
/* struct common data */
copy_v2_v2(AFD.dxt, dxt);
copy_v2_v2(AFD.dyt, dyt);
AFD.intpol = intpol;
AFD.extflag = extflag;
/* brecht: added stupid clamping here, large dx/dy can give very large
* filter sizes which take ages to render, it may be better to do this
* more intelligently later in the code .. probably it's not noticeable */
if (AFD.dxt[0]*AFD.dxt[0] + AFD.dxt[1]*AFD.dxt[1] > 2.0f*2.0f)
mul_v2_fl(AFD.dxt, 2.0f/len_v2(AFD.dxt));
if (AFD.dyt[0]*AFD.dyt[0] + AFD.dyt[1]*AFD.dyt[1] > 2.0f*2.0f)
mul_v2_fl(AFD.dyt, 2.0f/len_v2(AFD.dyt));
/* choice: */
if (tex->imaflag & TEX_MIPMAP) {
ImBuf *previbuf, *curibuf;
float levf;
int maxlev;
ImBuf *mipmaps[IMB_MIPMAP_LEVELS + 1];
/* modify ellipse minor axis if too eccentric, use for area sampling as well
* scaling dxt/dyt as done in pbrt is not the same
* (as in ewa_eval(), scale by sqrt(ibuf->x) to maximize precision) */
const float ff = sqrtf(ibuf->x), q = ibuf->y/ff;
const float Ux = dxt[0]*ff, Vx = dxt[1]*q, Uy = dyt[0]*ff, Vy = dyt[1]*q;
const float A = Vx*Vx + Vy*Vy;
const float B = -2.f*(Ux*Vx + Uy*Vy);
const float C = Ux*Ux + Uy*Uy;
const float F = A*C - B*B*0.25f;
float a, b, th, ecc;
BLI_ewa_imp2radangle(A, B, C, F, &a, &b, &th, &ecc);
if (tex->texfilter == TXF_FELINE) {
float fProbes;
a *= ff;
b *= ff;
a = max_ff(a, 1.0f);
b = max_ff(b, 1.0f);
fProbes = 2.f*(a / b) - 1.f;
AFD.iProbes = round_fl_to_int(fProbes);
AFD.iProbes = MIN2(AFD.iProbes, tex->afmax);
if (AFD.iProbes < fProbes)
b = 2.f*a / (float)(AFD.iProbes + 1);
AFD.majrad = a/ff;
AFD.minrad = b/ff;
AFD.theta = th;
AFD.dusc = 1.f/ff;
AFD.dvsc = ff / (float)ibuf->y;
}
else { /* EWA & area */
if (ecc > (float)tex->afmax) b = a / (float)tex->afmax;
b *= ff;
}
maxd = max_ff(b, 1e-8f);
levf = ((float)M_LOG2E) * logf(maxd);
curmap = 0;
maxlev = 1;
mipmaps[0] = ibuf;
while (curmap < IMB_MIPMAP_LEVELS) {
mipmaps[curmap + 1] = ibuf->mipmap[curmap];
if (ibuf->mipmap[curmap]) maxlev++;
curmap++;
}
/* mipmap level */
if (levf < 0.f) { /* original image only */
previbuf = curibuf = mipmaps[0];
levf = 0.f;
}
else if (levf >= maxlev - 1) {
previbuf = curibuf = mipmaps[maxlev - 1];
levf = 0.f;
if (tex->texfilter == TXF_FELINE) AFD.iProbes = 1;
}
else {
const int lev = isnan(levf) ? 0 : (int)levf;
curibuf = mipmaps[lev];
previbuf = mipmaps[lev + 1];
levf -= floorf(levf);
}
/* filter functions take care of interpolation themselves, no need to modify dxt/dyt here */
if (texres->nor && ((tex->imaflag & TEX_NORMALMAP) == 0)) {
/* color & normal */
filterfunc(texres, curibuf, fx, fy, &AFD);
val1 = texres->tr + texres->tg + texres->tb;
filterfunc(&texr, curibuf, fx + dxt[0], fy + dxt[1], &AFD);
val2 = texr.tr + texr.tg + texr.tb;
filterfunc(&texr, curibuf, fx + dyt[0], fy + dyt[1], &AFD);
val3 = texr.tr + texr.tg + texr.tb;
/* don't switch x or y! */
texres->nor[0] = val1 - val2;
texres->nor[1] = val1 - val3;
if (previbuf != curibuf) { /* interpolate */
filterfunc(&texr, previbuf, fx, fy, &AFD);
/* rgb */
texres->tr += levf*(texr.tr - texres->tr);
texres->tg += levf*(texr.tg - texres->tg);
texres->tb += levf*(texr.tb - texres->tb);
texres->ta += levf*(texr.ta - texres->ta);
/* normal */
val1 += levf*((texr.tr + texr.tg + texr.tb) - val1);
filterfunc(&texr, previbuf, fx + dxt[0], fy + dxt[1], &AFD);
val2 += levf*((texr.tr + texr.tg + texr.tb) - val2);
filterfunc(&texr, previbuf, fx + dyt[0], fy + dyt[1], &AFD);
val3 += levf*((texr.tr + texr.tg + texr.tb) - val3);
texres->nor[0] = val1 - val2; /* vals have been interpolated above! */
texres->nor[1] = val1 - val3;
}
}
else { /* color */
filterfunc(texres, curibuf, fx, fy, &AFD);
if (previbuf != curibuf) { /* interpolate */
filterfunc(&texr, previbuf, fx, fy, &AFD);
texres->tr += levf*(texr.tr - texres->tr);
texres->tg += levf*(texr.tg - texres->tg);
texres->tb += levf*(texr.tb - texres->tb);
texres->ta += levf*(texr.ta - texres->ta);
}
if (tex->texfilter != TXF_EWA) {
alpha_clip_aniso(ibuf, fx-minx, fy-miny, fx+minx, fy+miny, extflag, texres);
}
}
}
else { /* no mipmap */
/* filter functions take care of interpolation themselves, no need to modify dxt/dyt here */
if (tex->texfilter == TXF_FELINE) {
const float ff = sqrtf(ibuf->x), q = ibuf->y/ff;
const float Ux = dxt[0]*ff, Vx = dxt[1]*q, Uy = dyt[0]*ff, Vy = dyt[1]*q;
const float A = Vx*Vx + Vy*Vy;
const float B = -2.f*(Ux*Vx + Uy*Vy);
const float C = Ux*Ux + Uy*Uy;
const float F = A*C - B*B*0.25f;
float a, b, th, ecc, fProbes;
BLI_ewa_imp2radangle(A, B, C, F, &a, &b, &th, &ecc);
a *= ff;
b *= ff;
a = max_ff(a, 1.0f);
b = max_ff(b, 1.0f);
fProbes = 2.f*(a / b) - 1.f;
/* no limit to number of Probes here */
AFD.iProbes = round_fl_to_int(fProbes);
if (AFD.iProbes < fProbes) b = 2.f*a / (float)(AFD.iProbes + 1);
AFD.majrad = a/ff;
AFD.minrad = b/ff;
AFD.theta = th;
AFD.dusc = 1.f/ff;
AFD.dvsc = ff / (float)ibuf->y;
}
if (texres->nor && ((tex->imaflag & TEX_NORMALMAP) == 0)) {
/* color & normal */
filterfunc(texres, ibuf, fx, fy, &AFD);
val1 = texres->tr + texres->tg + texres->tb;
filterfunc(&texr, ibuf, fx + dxt[0], fy + dxt[1], &AFD);
val2 = texr.tr + texr.tg + texr.tb;
filterfunc(&texr, ibuf, fx + dyt[0], fy + dyt[1], &AFD);
val3 = texr.tr + texr.tg + texr.tb;
/* don't switch x or y! */
texres->nor[0] = val1 - val2;
texres->nor[1] = val1 - val3;
}
else {
filterfunc(texres, ibuf, fx, fy, &AFD);
if (tex->texfilter != TXF_EWA) {
alpha_clip_aniso(ibuf, fx-minx, fy-miny, fx+minx, fy+miny, extflag, texres);
}
}
}
if (tex->imaflag & TEX_CALCALPHA)
texres->ta = texres->tin = texres->ta * max_fff(texres->tr, texres->tg, texres->tb);
else
texres->tin = texres->ta;
if (tex->flag & TEX_NEGALPHA) texres->ta = 1.f - texres->ta;
if ((R.flag & R_SEC_FIELD) && (ibuf->flags & IB_fields))
ibuf->rect -= ibuf->x*ibuf->y;
if (texres->nor && (tex->imaflag & TEX_NORMALMAP)) { /* normal from color */
/* The invert of the red channel is to make
* the normal map compliant with the outside world.
* It needs to be done because in Blender
* the normal used in the renderer points inward. It is generated
* this way in calc_vertexnormals(). Should this ever change
* this negate must be removed. */
texres->nor[0] = -2.f*(texres->tr - 0.5f);
texres->nor[1] = 2.f*(texres->tg - 0.5f);
texres->nor[2] = 2.f*(texres->tb - 0.5f);
}
/* de-premul, this is being premulled in shade_input_do_shade()
* TXF: this currently does not (yet?) work properly, destroys edge AA in clip/checker mode, so for now commented out
* also disabled in imagewraposa() to be able to compare results with blender's default texture filtering */
/* brecht: tried to fix this, see "TXF alpha" comments */
/* do not de-premul for generated alpha, it is already in straight */
if (texres->ta!=1.0f && texres->ta>1e-4f && !(tex->imaflag & TEX_CALCALPHA)) {
fx = 1.f/texres->ta;
texres->tr *= fx;
texres->tg *= fx;
texres->tb *= fx;
}
if (ima)
BKE_image_pool_release_ibuf(ima, ibuf, pool);
BRICONTRGB;
return retval;
}
