bool
TextureSystemImpl::accum3d_sample_bilinear (const Imath::V3f &P, int miplevel,
TextureFile &texturefile,
PerThreadInfo *thread_info,
TextureOpt &options,
float weight, float *accum, float *daccumds,
float *daccumdt, float *daccumdr)
{
const ImageSpec &spec (texturefile.spec (options.subimage, miplevel));
const ImageCacheFile::LevelInfo &levelinfo (texturefile.levelinfo(options.subimage,miplevel));
// As passed in, (s,t) map the texture to (0,1). Remap to texel coords
// and subtract 0.5 because samples are at texel centers.
float s = P[0] * spec.full_width + spec.full_x - 0.5f;
float t = P[1] * spec.full_height + spec.full_y - 0.5f;
float r = P[2] * spec.full_depth + spec.full_z - 0.5f;
int sint, tint, rint;
float sfrac = floorfrac (s, &sint);
float tfrac = floorfrac (t, &tint);
float rfrac = floorfrac (r, &rint);
// Now (sint,tint,rint) are the integer coordinates of the texel to the
// immediate "upper left" of the lookup point, and (sfrac,tfrac,rfrac) are
// the amount that the lookup point is actually offset from the
// texel center (with (1,1) being all the way to the next texel down
// and to the right).
// Wrap
DASSERT (options.swrap_func != NULL && options.twrap_func != NULL &&
options.rwrap_func != NULL);
int stex[2], ttex[2], rtex[2]; // Texel coords
stex[0] = sint; stex[1] = sint+1;
ttex[0] = tint; ttex[1] = tint+1;
rtex[0] = rint; rtex[1] = rint+1;
// bool svalid[2], tvalid[2], rvalid[2]; // Valid texels? false means black border
union { bool bvalid[6]; unsigned long long ivalid; } valid_storage;
valid_storage.ivalid = 0;
DASSERT (sizeof(valid_storage) >= 6*sizeof(bool));
const unsigned long long none_valid = 0;
const unsigned long long all_valid = 0x010101010101LL;
DASSERT (__LITTLE_ENDIAN__ && "this trick won't work with big endian");
bool *svalid = valid_storage.bvalid;
bool *tvalid = valid_storage.bvalid + 2;
bool *rvalid = valid_storage.bvalid + 4;
svalid[0] = options.swrap_func (stex[0], spec.x, spec.width);
svalid[1] = options.swrap_func (stex[1], spec.x, spec.width);
tvalid[0] = options.twrap_func (ttex[0], spec.y, spec.height);
tvalid[1] = options.twrap_func (ttex[1], spec.y, spec.height);
rvalid[0] = options.rwrap_func (rtex[0], spec.z, spec.depth);
rvalid[1] = options.rwrap_func (rtex[1], spec.z, spec.depth);
// Account for crop windows
if (! levelinfo.full_pixel_range) {
svalid[0] &= (stex[0] >= spec.x && stex[0] < spec.x+spec.width);
svalid[1] &= (stex[1] >= spec.x && stex[1] < spec.x+spec.width);
tvalid[0] &= (ttex[0] >= spec.y && ttex[0] < spec.y+spec.height);
tvalid[1] &= (ttex[1] >= spec.y && ttex[1] < spec.y+spec.height);
rvalid[0] &= (rtex[0] >= spec.z && rtex[0] < spec.z+spec.depth);
rvalid[1] &= (rtex[1] >= spec.z && rtex[1] < spec.z+spec.depth);
}
// if (! (svalid[0] | svalid[1] | tvalid[0] | tvalid[1] | rvalid[0] | rvalid[1]))
if (valid_storage.ivalid == none_valid)
return true; // All texels we need were out of range and using 'black' wrap
int tilewidthmask = spec.tile_width - 1; // e.g. 63
int tileheightmask = spec.tile_height - 1;
int tiledepthmask = spec.tile_depth - 1;
const unsigned char *texel[2][2][2];
TileRef savetile[2][2][2];
static float black[8] = { 0, 0, 0, 0, 0, 0, 0, 0 };
int tile_s = (stex[0] - spec.x) % spec.tile_width;
int tile_t = (ttex[0] - spec.y) % spec.tile_height;
int tile_r = (rtex[0] - spec.z) % spec.tile_depth;
bool s_onetile = (tile_s != tilewidthmask) & (stex[0]+1 == stex[1]);
bool t_onetile = (tile_t != tileheightmask) & (ttex[0]+1 == ttex[1]);
bool r_onetile = (tile_r != tiledepthmask) & (rtex[0]+1 == rtex[1]);
bool onetile = (s_onetile & t_onetile & r_onetile);
size_t channelsize = texturefile.channelsize();
size_t pixelsize = texturefile.pixelsize();
if (onetile &&
valid_storage.ivalid == all_valid) {
// Shortcut if all the texels we need are on the same tile
TileID id (texturefile, options.subimage, miplevel,
stex[0] - tile_s, ttex[0] - tile_t, rtex[0] - tile_r);
bool ok = find_tile (id, thread_info);
if (! ok)
error ("%s", m_imagecache->geterror().c_str());
TileRef &tile (thread_info->tile);
if (! tile->valid())
return false;
size_t tilepel = (tile_r * spec.tile_height + tile_t) * spec.tile_width + tile_s;
size_t offset = (spec.nchannels * tilepel + options.firstchannel) * channelsize;
DASSERT ((size_t)offset < spec.tile_width*spec.tile_height*spec.tile_depth*pixelsize);
const unsigned char *b = tile->bytedata() + offset;
texel[0][0][0] = b;
texel[0][0][1] = b + pixelsize;
texel[0][1][0] = b + pixelsize * spec.tile_width;
texel[0][1][1] = b + pixelsize * spec.tile_width + pixelsize;
b += pixelsize * spec.tile_width * spec.tile_height;
texel[1][0][0] = b;
texel[1][0][1] = b + pixelsize;
texel[1][1][0] = b + pixelsize * spec.tile_width;
texel[1][1][1] = b + pixelsize * spec.tile_width + pixelsize;
} else {
for (int k = 0; k < 2; ++k) {
for (int j = 0; j < 2; ++j) {
for (int i = 0; i < 2; ++i) {
if (! (svalid[i] && tvalid[j] && rvalid[k])) {
texel[k][j][i] = (unsigned char *)black;
continue;
}
tile_s = (stex[i] - spec.x) % spec.tile_width;
tile_t = (ttex[j] - spec.y) % spec.tile_height;
tile_r = (rtex[k] - spec.z) % spec.tile_depth;
TileID id (texturefile, options.subimage, miplevel,
stex[i] - tile_s, ttex[j] - tile_t,
rtex[k] - tile_r);
bool ok = find_tile (id, thread_info);
if (! ok)
error ("%s", m_imagecache->geterror().c_str());
TileRef &tile (thread_info->tile);
if (! tile->valid())
return false;
savetile[k][j][i] = tile;
size_t tilepel = (tile_r * spec.tile_height + tile_t) * spec.tile_width + tile_s;
size_t offset = (spec.nchannels * tilepel + options.firstchannel) * channelsize;
#if DEBUG
if ((size_t)offset >= spec.tile_width*spec.tile_height*spec.tile_depth*pixelsize)
std::cerr << "offset=" << offset << ", whd " << spec.tile_width << ' ' << spec.tile_height << ' ' << spec.tile_depth << " pixsize " << pixelsize << "\n";
#endif
DASSERT ((size_t)offset < spec.tile_width*spec.tile_height*spec.tile_depth*pixelsize);
texel[k][j][i] = tile->bytedata() + offset;
DASSERT (tile->id() == id);
}
}
}
}
// FIXME -- optimize the above loop by unrolling
if (channelsize == 1) {
// special case for 8-bit tiles
int c;
for (c = 0; c < options.actualchannels; ++c)
accum[c] += weight * trilerp (uchar2float(texel[0][0][0][c]), uchar2float(texel[0][0][1][c]),
uchar2float(texel[0][1][0][c]), uchar2float(texel[0][1][1][c]),
uchar2float(texel[1][0][0][c]), uchar2float(texel[1][0][1][c]),
uchar2float(texel[1][1][0][c]), uchar2float(texel[1][1][1][c]),
sfrac, tfrac, rfrac);
if (daccumds) {
float scalex = weight * spec.full_width;
float scaley = weight * spec.full_height;
float scalez = weight * spec.full_depth;
for (c = 0; c < options.actualchannels; ++c) {
daccumds[c] += scalex * bilerp(
uchar2float(texel[0][0][1][c]) - uchar2float(texel[0][0][0][c]),
uchar2float(texel[0][1][1][c]) - uchar2float(texel[0][1][0][c]),
uchar2float(texel[1][0][1][c]) - uchar2float(texel[1][0][0][c]),
uchar2float(texel[1][1][1][c]) - uchar2float(texel[1][1][0][c]),
tfrac, rfrac
);
daccumdt[c] += scaley * bilerp(
uchar2float(texel[0][1][0][c]) - uchar2float(texel[0][0][0][c]),
uchar2float(texel[0][1][1][c]) - uchar2float(texel[0][0][1][c]),
uchar2float(texel[1][1][0][c]) - uchar2float(texel[1][0][0][c]),
uchar2float(texel[1][1][1][c]) - uchar2float(texel[1][0][1][c]),
sfrac, rfrac
);
daccumdr[c] += scalez * bilerp(
uchar2float(texel[0][1][0][c]) - uchar2float(texel[1][1][0][c]),
uchar2float(texel[0][1][1][c]) - uchar2float(texel[1][1][1][c]),
uchar2float(texel[0][0][1][c]) - uchar2float(texel[1][0][0][c]),
uchar2float(texel[0][1][1][c]) - uchar2float(texel[1][1][1][c]),
sfrac, tfrac
);
}
}
} else {
// General case for float tiles
trilerp_mad ((const float *)texel[0][0][0], (const float *)texel[0][0][1],
(const float *)texel[0][1][0], (const float *)texel[0][1][1],
(const float *)texel[1][0][0], (const float *)texel[1][0][1],
(const float *)texel[1][1][0], (const float *)texel[1][1][1],
sfrac, tfrac, rfrac, weight, options.actualchannels, accum);
if (daccumds) {
float scalex = weight * spec.full_width;
float scaley = weight * spec.full_height;
float scalez = weight * spec.full_depth;
for (int c = 0; c < options.actualchannels; ++c) {
daccumds[c] += scalex * bilerp(
((const float *) texel[0][0][1])[c] - ((const float *) texel[0][0][0])[c],
((const float *) texel[0][1][1])[c] - ((const float *) texel[0][1][0])[c],
((const float *) texel[1][0][1])[c] - ((const float *) texel[1][0][0])[c],
((const float *) texel[1][1][1])[c] - ((const float *) texel[1][1][0])[c],
tfrac, rfrac
);
daccumdt[c] += scaley * bilerp(
((const float *) texel[0][1][0])[c] - ((const float *) texel[0][0][0])[c],
((const float *) texel[0][1][1])[c] - ((const float *) texel[0][0][1])[c],
((const float *) texel[1][1][0])[c] - ((const float *) texel[1][0][0])[c],
((const float *) texel[1][1][1])[c] - ((const float *) texel[1][0][1])[c],
sfrac, rfrac
);
daccumdr[c] += scalez * bilerp(
((const float *) texel[0][1][0])[c] - ((const float *) texel[1][1][0])[c],
((const float *) texel[0][1][1])[c] - ((const float *) texel[1][1][1])[c],
((const float *) texel[0][0][1])[c] - ((const float *) texel[1][0][0])[c],
((const float *) texel[0][1][1])[c] - ((const float *) texel[1][1][1])[c],
sfrac, tfrac
);
}
}
}
return true;
}
bool
TextureSystemImpl::environment (TextureHandle *texture_handle_,
Perthread *thread_info_,
TextureOpt &options, const Imath::V3f &_R,
const Imath::V3f &_dRdx, const Imath::V3f &_dRdy,
float *result)
{
PerThreadInfo *thread_info = (PerThreadInfo *)thread_info_;
TextureFile *texturefile = (TextureFile *)texture_handle_;
ImageCacheStatistics &stats (thread_info->m_stats);
++stats.environment_batches;
++stats.environment_queries;
if (! texturefile || texturefile->broken())
return missing_texture (options, result);
const ImageSpec &spec (texturefile->spec(options.subimage, 0));
options.swrap_func = texturefile->m_sample_border ?
wrap_periodic_sharedborder : wrap_periodic;
options.twrap_func = wrap_clamp;
options.envlayout = LayoutLatLong;
int actualchannels = Imath::clamp (spec.nchannels - options.firstchannel,
0, options.nchannels);
options.actualchannels = actualchannels;
// Initialize results to 0. We'll add from here on as we sample.
float* dresultds = options.dresultds;
float* dresultdt = options.dresultdt;
for (int c = 0; c < options.actualchannels; ++c) {
result[c] = 0;
if (dresultds) dresultds[c] = 0;
if (dresultdt) dresultdt[c] = 0;
}
// If the user only provided us with one pointer, clear both to simplify
// the rest of the code, but only after we zero out the data for them so
// they know something went wrong.
if (!(dresultds && dresultdt))
dresultds = dresultdt = NULL;
// Calculate unit-length vectors in the direction of R, R+dRdx, R+dRdy.
// These define the ellipse we're filtering over.
Imath::V3f R = _R; R.normalize(); // center
Imath::V3f Rx = _R + _dRdx; Rx.normalize(); // x axis of the ellipse
Imath::V3f Ry = _R + _dRdy; Ry.normalize(); // y axis of the ellipse
// angles formed by the ellipse axes.
float xfilt_noblur = std::max (safe_acosf(R.dot(Rx)), 1e-8f);
float yfilt_noblur = std::max (safe_acosf(R.dot(Ry)), 1e-8f);
int naturalres = int((float)M_PI / std::min (xfilt_noblur, yfilt_noblur));
// FIXME -- figure naturalres sepearately for s and t
// FIXME -- ick, why is it x and y at all, shouldn't it be s and t?
// N.B. naturalres formulated for latlong
// Account for width and blur
float xfilt = xfilt_noblur * options.swidth + options.sblur;
float yfilt = yfilt_noblur * options.twidth + options.tblur;
// Figure out major versus minor, and aspect ratio
Imath::V3f Rmajor; // major axis
float majorlength, minorlength;
bool x_is_majoraxis = (xfilt >= yfilt);
if (x_is_majoraxis) {
Rmajor = Rx;
majorlength = xfilt;
minorlength = yfilt;
} else {
Rmajor = Ry;
majorlength = yfilt;
minorlength = xfilt;
}
accum_prototype accumer;
long long *probecount;
switch (options.interpmode) {
case TextureOpt::InterpClosest :
accumer = &TextureSystemImpl::accum_sample_closest;
probecount = &stats.closest_interps;
break;
case TextureOpt::InterpBilinear :
accumer = &TextureSystemImpl::accum_sample_bilinear;
probecount = &stats.bilinear_interps;
break;
case TextureOpt::InterpBicubic :
accumer = &TextureSystemImpl::accum_sample_bicubic;
probecount = &stats.cubic_interps;
break;
default:
accumer = NULL;
probecount = NULL;
break;
}
TextureOpt::MipMode mipmode = options.mipmode;
bool aniso = (mipmode == TextureOpt::MipModeDefault ||
mipmode == TextureOpt::MipModeAniso);
float aspect, trueaspect, filtwidth;
int nsamples;
float invsamples;
if (aniso) {
aspect = anisotropic_aspect (majorlength, minorlength, options, trueaspect);
filtwidth = minorlength;
if (trueaspect > stats.max_aniso)
stats.max_aniso = trueaspect;
nsamples = std::max (1, (int) ceilf (aspect - 0.25f));
invsamples = 1.0f / nsamples;
} else {
filtwidth = options.conservative_filter ? majorlength : minorlength;
nsamples = 1;
invsamples = 1.0f;
}
ImageCacheFile::SubimageInfo &subinfo (texturefile->subimageinfo(options.subimage));
// FIXME -- assuming latlong
bool ok = true;
float pos = -0.5f + 0.5f * invsamples;
for (int sample = 0; sample < nsamples; ++sample, pos += invsamples) {
Imath::V3f Rsamp = R + pos*Rmajor;
float s, t;
vector_to_latlong (Rsamp, texturefile->m_y_up, s, t);
// Determine the MIP-map level(s) we need: we will blend
// data(miplevel[0]) * (1-levelblend) + data(miplevel[1]) * levelblend
int miplevel[2] = { -1, -1 };
float levelblend = 0;
int nmiplevels = (int)subinfo.levels.size();
for (int m = 0; m < nmiplevels; ++m) {
// Compute the filter size in raster space at this MIP level.
// Filters are in radians, and the vertical resolution of a
// latlong map is PI radians. So to compute the raster size of
// our filter width...
float filtwidth_ras = subinfo.spec(m).full_height * filtwidth * M_1_PI;
// Once the filter width is smaller than one texel at this level,
// we've gone too far, so we know that we want to interpolate the
// previous level and the current level. Note that filtwidth_ras
// is expected to be >= 0.5, or would have stopped one level ago.
if (filtwidth_ras <= 1) {
miplevel[0] = m-1;
miplevel[1] = m;
levelblend = Imath::clamp (2.0f - 1.0f/filtwidth_ras, 0.0f, 1.0f);
break;
}
}
if (miplevel[1] < 0) {
// We'd like to blur even more, but make due with the coarsest
// MIP level.
miplevel[0] = nmiplevels - 1;
miplevel[1] = miplevel[0];
levelblend = 0;
} else if (miplevel[0] < 0) {
// We wish we had even more resolution than the finest MIP level,
// but tough for us.
miplevel[0] = 0;
miplevel[1] = 0;
levelblend = 0;
}
if (options.mipmode == TextureOpt::MipModeOneLevel) {
// Force use of just one mipmap level
miplevel[1] = miplevel[0];
levelblend = 0;
} else if (mipmode == TextureOpt::MipModeNoMIP) {
// Just sample from lowest level
miplevel[0] = 0;
miplevel[1] = 0;
levelblend = 0;
}
float levelweight[2] = { 1.0f - levelblend, levelblend };
int npointson = 0;
for (int level = 0; level < 2; ++level) {
if (! levelweight[level])
continue;
++npointson;
int lev = miplevel[level];
if (options.interpmode == TextureOpt::InterpSmartBicubic) {
if (lev == 0 ||
(texturefile->spec(options.subimage,lev).full_height < naturalres/2)) {
accumer = &TextureSystemImpl::accum_sample_bicubic;
++stats.cubic_interps;
} else {
accumer = &TextureSystemImpl::accum_sample_bilinear;
++stats.bilinear_interps;
}
} else {
*probecount += 1;
}
ok &= (this->*accumer) (s, t, miplevel[level], *texturefile,
thread_info, options,
levelweight[level]*invsamples, result,
dresultds, dresultdt);
}
}
stats.aniso_probes += nsamples;
++stats.aniso_queries;
if (actualchannels < options.nchannels)
fill_channels (spec, options, result);
return ok;
}
bool
TextureSystemImpl::accum3d_sample_closest (const Imath::V3f &P, int miplevel,
TextureFile &texturefile,
PerThreadInfo *thread_info,
TextureOpt &options,
float weight, float *accum, float *daccumds,
float *daccumdt, float *daccumdr)
{
const ImageSpec &spec (texturefile.spec (options.subimage, miplevel));
const ImageCacheFile::LevelInfo &levelinfo (texturefile.levelinfo(options.subimage,miplevel));
// As passed in, (s,t) map the texture to (0,1). Remap to texel coords.
float s = P[0] * spec.full_width + spec.full_x;
float t = P[1] * spec.full_height + spec.full_y;
float r = P[2] * spec.full_depth + spec.full_z;
int stex, ttex, rtex; // Texel coordintes
(void) floorfrac (s, &stex); // don't need fractional result
(void) floorfrac (t, &ttex);
(void) floorfrac (r, &rtex);
// Wrap
DASSERT (options.swrap_func != NULL && options.twrap_func != NULL &&
options.rwrap_func != NULL);
bool svalid, tvalid, rvalid; // Valid texels? false means black border
svalid = options.swrap_func (stex, spec.x, spec.width);
tvalid = options.twrap_func (ttex, spec.y, spec.height);
rvalid = options.rwrap_func (rtex, spec.z, spec.depth);
if (! levelinfo.full_pixel_range) {
svalid &= (stex >= spec.x && stex < (spec.x+spec.width)); // data window
tvalid &= (ttex >= spec.y && ttex < (spec.y+spec.height));
rvalid &= (rtex >= spec.z && rtex < (spec.z+spec.depth));
}
if (! (svalid & tvalid & rvalid)) {
// All texels we need were out of range and using 'black' wrap.
return true;
}
int tile_s = (stex - spec.x) % spec.tile_width;
int tile_t = (ttex - spec.y) % spec.tile_height;
int tile_r = (rtex - spec.z) % spec.tile_depth;
TileID id (texturefile, options.subimage, miplevel,
stex - tile_s, ttex - tile_t, rtex - tile_r);
bool ok = find_tile (id, thread_info);
if (! ok)
error ("%s", m_imagecache->geterror().c_str());
TileRef &tile (thread_info->tile);
if (! tile || ! ok)
return false;
size_t channelsize = texturefile.channelsize();
int tilepel = (tile_r * spec.tile_height + tile_t) * spec.tile_width + tile_s;
int offset = spec.nchannels * tilepel + options.firstchannel;
DASSERT ((size_t)offset < spec.nchannels*spec.tile_pixels());
if (channelsize == 1) {
// special case for 8-bit tiles
const unsigned char *texel = tile->bytedata() + offset;
for (int c = 0; c < options.actualchannels; ++c)
accum[c] += weight * uchar2float(texel[c]);
} else {
// General case for float tiles
const float *texel = tile->data() + offset;
for (int c = 0; c < options.actualchannels; ++c)
accum[c] += weight * texel[c];
}
return true;
}