T* doPermute(T* pIn, const std::vector<int>& dimsVect)
{
int iDims = pIn->getDims();
int* piDimsArray = pIn->getDimsArray();
int* piOffset = new int[iDims];
int* piMaxOffset = new int[iDims];
int* piIndex = new int[iDims]();
computeOffsets(iDims, piDimsArray, dimsVect, piOffset, piMaxOffset);
T* pOut = pIn->clone();
for (int iSource = 0, iDest = 0; iSource < pIn->getSize(); iSource++)
{
pOut->set(iDest, pIn->get(iSource));
for (int j = 0; j < iDims; j++)
{
++piIndex[j];
iDest += piOffset[j];
if (piIndex[j] < piDimsArray[j])
{
break;
}
iDest -= piMaxOffset[j];
piIndex[j] = 0;
}
}
delete[] piIndex;
delete[] piOffset;
delete[] piMaxOffset;
return pOut;
}
void PtexReader::readFaceData(FilePos pos, FaceDataHeader fdh, Res res, int levelid,
FaceData*& face)
{
// keep new face local until fully initialized
FaceData* volatile newface = 0;
seek(pos);
switch (fdh.encoding()) {
case enc_constant:
{
ConstantFace* pf = new ConstantFace((void**)&face, _cache, _pixelsize);
readBlock(pf->data(), _pixelsize);
if (levelid==0 && _premultiply && _header.hasAlpha())
PtexUtils::multalpha(pf->data(), 1, _header.datatype,
_header.nchannels, _header.alphachan);
newface = pf;
}
break;
case enc_tiled:
{
Res tileres;
readBlock(&tileres, sizeof(tileres));
uint32_t tileheadersize;
readBlock(&tileheadersize, sizeof(tileheadersize));
TiledFace* tf = new TiledFace((void**)&face, _cache, res, tileres, levelid, this);
readZipBlock(&tf->_fdh[0], tileheadersize, FaceDataHeaderSize * tf->_ntiles);
computeOffsets(tell(), tf->_ntiles, &tf->_fdh[0], &tf->_offsets[0]);
newface = tf;
}
break;
case enc_zipped:
case enc_diffzipped:
{
int uw = res.u(), vw = res.v();
int npixels = uw * vw;
int unpackedSize = _pixelsize * npixels;
PackedFace* pf = new PackedFace((void**)&face, _cache,
res, _pixelsize, unpackedSize);
bool useMalloc = unpackedSize > AllocaMax;
void* tmp = useMalloc ? malloc(unpackedSize) : alloca(unpackedSize);
readZipBlock(tmp, fdh.blocksize(), unpackedSize);
if (fdh.encoding() == enc_diffzipped)
PtexUtils::decodeDifference(tmp, unpackedSize, _header.datatype);
PtexUtils::interleave(tmp, uw * DataSize(_header.datatype), uw, vw,
pf->data(), uw * _pixelsize,
_header.datatype, _header.nchannels);
if (levelid==0 && _premultiply && _header.hasAlpha())
PtexUtils::multalpha(pf->data(), npixels, _header.datatype,
_header.nchannels, _header.alphachan);
newface = pf;
if (useMalloc) free(tmp);
}
break;
}
face = newface;
}
/*private*/
void
OffsetPointGenerator::extractPoints(const LineString* line)
{
const CoordinateSequence& pts = *(line->getCoordinatesRO());
assert(pts.size() > 1 );
for (size_t i=0, n=pts.size()-1; i<n; ++i)
{
computeOffsets(pts[i], pts[i + 1]);
}
}
esvmArr2_f *convolve2D(float im[], const int irows, const int icols, float kern[], const int krows, const int kcols)
{
esvmArr2_f *output = (esvmArr2_f *)esvmMalloc(sizeof(esvmArr2_f));
output->arr = (float *)esvmCalloc(irows*icols,sizeof(float));
int *offset = (int *)esvmMalloc(krows*kcols*sizeof(int));
computeOffsets(offset,krows,kcols,irows,icols);
ispc::convolve2D(im, irows, icols, kern, krows, kcols,(ispc::esvmArr2_f *)output,offset);
return output;
}
EBStenVarCoef::
EBStenVarCoef(const Vector<VolIndex>& a_srcVofs,
const BaseIVFAB<VoFStencil>& a_vofStencil,
const Box& a_box,
const EBISBox& a_ebisBox,
const IntVect& a_ghostVect,
int a_varDest)
: m_box( a_box ),
m_ebisBox( a_ebisBox ),
m_ghostVect( a_ghostVect ),
m_destVar( a_varDest )
{
CH_TIME("EBStenVarCoef::EBStenVarCoef");
computeOffsets(a_srcVofs, a_vofStencil);
}
void PtexReader::readLevel(int levelid, Level*& level)
{
// temporarily release cache lock so other threads can proceed
_cache->cachelock.unlock();
// get read lock and make sure we still need to read
AutoMutex locker(readlock);
if (level) {
// another thread must have read it while we were waiting
_cache->cachelock.lock();
// make sure it's still there now that we have the lock
if (level) {
level->ref();
return;
}
_cache->cachelock.unlock();
}
// go ahead and read the level
LevelInfo& l = _levelinfo[levelid];
// keep new level local until finished
Level* volatile newlevel = new Level((void**)&level, _cache, l.nfaces);
seek(_levelpos[levelid]);
readZipBlock(&newlevel->fdh[0], l.levelheadersize, FaceDataHeaderSize * l.nfaces);
computeOffsets(tell(), l.nfaces, &newlevel->fdh[0], &newlevel->offsets[0]);
// apply edits (if any) to level 0
if (levelid == 0) {
for (size_t i = 0, size = _faceedits.size(); i < size; i++) {
FaceEdit& e = _faceedits[i];
newlevel->fdh[e.faceid] = e.fdh;
newlevel->offsets[e.faceid] = e.pos;
}
}
// don't assign to result until level data is fully initialized
_cache->cachelock.lock();
level = newlevel;
// clean up unused data
_cache->purgeData();
}
esvmArr2_f *convolvePyramids(const esvmHogPyr *feats, const esvmHogPyr *whogs, const bool enablePadding,const int userTasks)
{
esvmArr2_i *offsets = (esvmArr2_i *)esvmMalloc(sizeof(esvmArr2_i)*feats->num*whogs->num);
esvmArr2_f *outputs = (esvmArr2_f *)esvmMalloc(sizeof(esvmArr2_f)*feats->num*whogs->num);
#pragma omp parallel for
for(int i=0;i<feats->num;i++) {
for(int j=0;j<whogs->num;j++) {
const int frows = feats->hogs[i]->rows;
const int fcols = feats->hogs[i]->cols;
const int fbins = feats->hogs[i]->bins;
const int wrows = whogs->hogs[j]->rows;
const int wcols = whogs->hogs[j]->cols;
offsets[i*whogs->num+j].arr = (int *)esvmMalloc(wrows*wcols*sizeof(int));
offsets[i*whogs->num+j].rows = wrows;
offsets[i*whogs->num+j].cols = wcols;
outputs[i*whogs->num+j].arr = (float *)esvmMalloc(frows*fcols*fbins*sizeof(float));
outputs[i*whogs->num+j].rows = frows;
outputs[i*whogs->num+j].cols = fcols;
computeOffsets((offsets[i*whogs->num+j].arr),wrows,wcols,frows,fcols);
}
}
ispc::convolvePyramids((ispc::esvmHogPyr *)feats,(ispc::esvmHogPyr *)whogs,(ispc::esvmArr2_i *)offsets,enablePadding,(ispc::esvmArr2_f *)outputs,userTasks);
if(__unlikely(enablePadding==true)) {
#pragma omp parallel for
for(int i=0;i<feats->num;i++) {
for(int j=0;j<whogs->num;j++) {
const int frows = feats->hogs[i]->rows;
const int fcols = feats->hogs[i]->cols;
outputs[i*whogs->num+j].arr = (float *)realloc((void *)outputs[i*whogs->num+j].arr,frows*fcols*sizeof(float));
}
}
}
else {
#pragma omp parallel for
for(int i=0;i<feats->num;i++) {
for(int j=0;j<whogs->num;j++) {
const int frows = feats->hogs[i]->rows;
const int fcols = feats->hogs[i]->cols;
const int wrows = whogs->hogs[j]->rows;
const int wcols = whogs->hogs[j]->cols;
const int apronRows = (wrows/2);
const int apronCols = (wcols/2);
const int rowEnd = frows-apronRows+(1^(wrows&1));
const int colEnd = fcols - apronCols+(1^(wcols&1));
const int urows = rowEnd-apronRows;
const int ucols = colEnd-apronCols;
float *uout = (float *)esvmMalloc(urows*ucols*sizeof(float));
float *output = outputs[i*whogs->num+j].arr;
for(int ii=0;ii<urows;ii++) {
float *tmp = uout+ii*ucols;
float *tmpout = output+(ii+apronRows)*fcols+apronCols;
for(int jj=0;jj<ucols;jj++) {
(*tmp++) = (*tmpout++);
}
}
free(output);
outputs[i*whogs->num+j].arr = uout;
outputs[i*whogs->num+j].rows = urows;
outputs[i*whogs->num+j].cols = ucols;
}
}
}
return outputs;
}
esvmArr2_f *ompConvolve3D(const esvmHog *feat, const esvmHog *whog, const bool enablePadding, const int userTasks)
{
if(__unlikely(feat==NULL||whog==NULL)) {
fprintf(stderr,"convolve3D:: got NULL features or NULL weights\n");
return NULL;
}
const int frows = feat->rows;
const int fcols = feat->cols;
const int fbins = feat->bins;
const int wrows = whog->rows;
const int wcols = whog->cols;
const int wbins = whog->bins;
if(__unlikely(wbins!=fbins || frows<wrows || fcols<wcols)) {
fprintf(stderr,"convolve3D:: dimensions of feature and weights don't match. feat(%d,%d,%d) ; weight (%d,%d,%d)\n",frows,fcols,fbins,wrows,wcols,wbins);
return NULL;
}
const int apronRows = floor(wrows/2);
const int apronCols = floor(wcols/2);
float *kern = whog->feature;
float *output = (float *)esvmCalloc(frows*fcols*fbins,sizeof(float));
if(__unlikely(output==NULL)) {
fprintf(stderr,"convolve3D:: Not enough memory for output array. Needed %ld bytes\n",frows*fcols*fbins*sizeof(float));
return NULL;
}
int *offset = (int *)esvmMalloc(wrows*wcols*sizeof(int));
if(__unlikely(offset==NULL)) {
fprintf(stderr,"convolve3D:: Not enough memory for offset array. Needed %ld bytes\n",wrows*wcols*sizeof(int));
return NULL;
}
computeOffsets(offset,wrows,wcols,frows,fcols);
const int dim1 = frows*fcols;
for(int i=0;i<fbins;i++) {
conv2DValid(feat->feature+i*dim1, frows, fcols, kern+i*wrows*wcols, wrows, wcols, apronRows, apronCols, offset, output+i*dim1, frows, fcols,1,0);
}
mergeConv(output, frows, fcols, fbins);
//cleanup
free(offset);
if(enablePadding==true) {
//shrink the output
output= (float *)realloc((void *)output,frows*fcols*sizeof(float));
esvmArr2_f *soln = (esvmArr2_f *)esvmMalloc(sizeof(esvmArr2_f));
soln->arr = output;
soln->rows = frows;
soln->cols = fcols;
return soln;
}
else {
const int rowEnd = frows-apronRows+(1^(wrows&1));
const int colEnd = fcols - apronCols+(1^(wcols&1));
const int urows = rowEnd-apronRows;
const int ucols = colEnd-apronCols;
float *uout = (float *)esvmMalloc(urows*ucols*sizeof(float));
for(int i=0;i<urows;i++) {
float *tmp = uout+i*ucols;
float *tmpout = output+(i+apronRows)*fcols+apronCols;
for(int j=0;j<ucols;j++) {
(*tmp++) = (*tmpout++);
}
}
free(output);
esvmArr2_f *soln = (esvmArr2_f *)esvmMalloc(sizeof(esvmArr2_f));
soln->arr = uout;
soln->rows = urows;
soln->cols = ucols;
return soln;
}
}
void Builder::fixup() {
computeOffsets();
fixupStringPoolOffsets();
fixupNameGroupPoolOffsets();
fixupMiscellaneousOffsets();
}
T *doNativePermute(T *pIn, const std::vector<int>& dimsVect)
{
int iDims = pIn->getDims();
int* piDimsArray = pIn->getDimsArray();
int* piIndex = new int[iDims]();
int* piOffset = new int[iDims];
int* piMaxOffset = new int[iDims];
computeOffsets(iDims, piDimsArray, dimsVect, piOffset, piMaxOffset);
T* pOut = pIn->clone();
typename T::type* pout = pOut->get();
if (pIn->isComplex())
{
typename T::type* poutImg = pOut->getImg();
for (typename T::type *pin = pIn->get(), *pinImg = pIn->getImg(); pin < pIn->get() + pIn->getSize(); pin++, pinImg++)
{
*pout = *pin;
*poutImg = *pinImg;
for (int j = 0; j < iDims; j++)
{
++piIndex[j];
pout += piOffset[j];
poutImg += piOffset[j];
if (piIndex[j] < piDimsArray[j])
{
break;
}
pout -= piMaxOffset[j];
poutImg -= piMaxOffset[j];
piIndex[j] = 0;
}
}
}
else
{
for (typename T::type *pin = pIn->get(); pin < pIn->get() + pIn->getSize(); pin++)
{
*pout = *pin;
for (int j = 0; j < iDims; j++)
{
++piIndex[j];
pout += piOffset[j];
if (piIndex[j] < piDimsArray[j])
{
break;
}
pout -= piMaxOffset[j];
piIndex[j] = 0;
}
}
}
delete[] piIndex;
delete[] piOffset;
delete[] piMaxOffset;
return pOut;
}