void pull( Treap * a ){
a->sum = Sum( a->l ) + Sum( a->r ) + a->val;
a->lsum = Sum( a->l ) + a->val + max( 0 , lSum( a->r ) );
if( a->l ) a->lsum = max( lSum( a->l ) , a->lsum );
a->rsum = Sum( a->r ) + a->val + max( 0 , rSum( a->l ) );
if( a->r ) a->rsum = max( rSum( a->r ) , a->rsum );
a->maxsum = max( 0 , rSum( a->l ) ) + a->val + max( 0 , lSum( a->r ) );
a->maxsum = max( a->maxsum , max( maxSum( a->l ) , maxSum( a->r ) ) );
a->sz = Size( a->l ) + Size( a->r ) + 1;
}
void hoSPIRITOperator<T>::sum_over_src_channel(const ARRAY_TYPE& x, ARRAY_TYPE& r)
{
try
{
boost::shared_ptr< std::vector<size_t> > dim = x.get_dimensions();
size_t NDim = dim->size();
if (NDim < 2) return;
std::vector<size_t> dimR(NDim - 1);
std::vector<size_t> dimRInternal = *dim;
dimRInternal[NDim - 2] = 1;
size_t d;
for (d = 0; d<NDim - 2; d++)
{
dimR[d] = (*dim)[d];
}
dimR[NDim - 2] = (*dim)[NDim - 1];
if (!r.dimensions_equal(&dimR))
{
r.create(&dimR);
}
if (x.get_size(NDim - 2) <= 1)
{
memcpy(r.begin(), x.begin(), x.get_number_of_bytes());
return;
}
hoNDArray<T> rSum(dimRInternal, r.begin());
GADGET_CATCH_THROW(Gadgetron::sum_over_dimension(x, rSum, NDim - 2));
}
catch (...)
{
GADGET_THROW("Errors in hoSPIRITOperator<T>::sum_over_src_channel(const ARRAY_TYPE& x, ARRAY_TYPE& r) ... ");
}
}
/*! \brief Tests the **reduce_sum** kernel.
* \details The kernel computes the sum of the elements of each row in an array.
*/
TEST (Reduce, reduce_sum)
{
try
{
const unsigned int rows = 1024;
const unsigned int cols = 1024;
const unsigned int bufferInSize = cols * rows * sizeof (cl_float);
const unsigned int bufferOutSize = rows * sizeof (cl_float);
// Setup the OpenCL environment
clutils::CLEnv clEnv;
clEnv.addContext (0);
clEnv.addQueue (0, 0, CL_QUEUE_PROFILING_ENABLE);
clEnv.addProgram (0, kernel_filename_reduce);
// Configure kernel execution parameters
clutils::CLEnvInfo<1> info (0, 0, 0, { 0 }, 0);
cl_algo::ICP::Reduce<cl_algo::ICP::ReduceConfig::SUM, cl_float> rSum (clEnv, info);
rSum.init (cols, rows);
// Initialize data (writes on staging buffer directly)
std::generate (rSum.hPtrIn, rSum.hPtrIn + bufferInSize / sizeof (cl_float), ICP::rNum_R_0_1);
// ICP::printBuffer ("Original:", rSum.hPtrIn, cols, rows);
rSum.write (); // Copy data to device
rSum.run (); // Execute kernels (~ 44 us)
cl_float *results = (cl_float *) rSum.read (); // Copy results to host
// ICP::printBuffer ("Received:", results, 1, rows);
// Produce reference array of distances
cl_float *refSum = new cl_float[rows];
ICP::cpuReduceSum (rSum.hPtrIn, refSum, cols, rows);
// ICP::printBuffer ("Expected:", refSum, 1, rows);
// Verify blurred output
float eps = 42000 * std::numeric_limits<float>::epsilon (); // 0.00500679
for (uint i = 0; i < rows; ++i)
ASSERT_LT (std::abs (refSum[i] - results[i]), eps);
// Profiling ===========================================================
if (profiling)
{
const int nRepeat = 1; /* Number of times to perform the tests. */
// CPU
clutils::CPUTimer<double, std::milli> cTimer;
clutils::ProfilingInfo<nRepeat> pCPU ("CPU");
for (int i = 0; i < nRepeat; ++i)
{
cTimer.start ();
ICP::cpuReduceSum (rSum.hPtrIn, refSum, cols, rows);
pCPU[i] = cTimer.stop ();
}
// GPU
clutils::GPUTimer<std::milli> gTimer (clEnv.devices[0][0]);
clutils::ProfilingInfo<nRepeat> pGPU ("GPU");
for (int i = 0; i < nRepeat; ++i)
pGPU[i] = rSum.run (gTimer);
// Benchmark
pGPU.print (pCPU, "Reduce<Sum>");
}
}
catch (const cl::Error &error)
{
std::cerr << error.what ()
<< " (" << clutils::getOpenCLErrorCodeString (error.err ())
<< ")" << std::endl;
exit (EXIT_FAILURE);
}
}
/* $begin rsum-c */
int rSum(int *Start, int Count)
{
if (Count <= 0)
return 0;
return *Start + rSum(Start+1, Count-1);
}
int rSum(int* start,int count){
if(count<=0)return 0;
return *start + rSum(start+1,count-1);
}
int rSumFits(int n, int m){
return rSum(n) <= m;
}
void CrossNLMFilter::Apply(const TwoDArray<Color> &img,
const TwoDArray<Feature> &featureImg,
const TwoDArray<Feature> &featureVarImg,
const TwoDArray<Color> &rImg,
const TwoDArray<Color> &varImg,
vector<TwoDArray<Color> > &fltArray,
vector<TwoDArray<float> > &mseArray,
vector<TwoDArray<float> > &priArray) const {
#pragma omp parallel for num_threads(PbrtOptions.nCores) schedule(static)
for(int taskId = 0; taskId < nTasks; taskId++) {
int txs, txe, tys, tye;
ComputeSubWindow(taskId, nTasks, width, height,
&txs, &txe, &tys, &tye);
for(int y = tys; y < tye; y++)
for(int x = txs; x < txe; x++) {
vector<Color> sum(scaleR.size(), Color(0.f));
vector<Color> rSum(scaleR.size(), Color(0.f));
vector<Color> rSumSq(scaleR.size(), Color(0.f));
vector<float> wSum(scaleR.size(), 0.f);
vector<vector<float> > wArray(scaleR.size());
for(size_t p = 0; p < wArray.size(); p++) {
wArray[p].resize(patchWidth*patchWidth);
}
Feature feature = featureImg(x, y);
Feature featureVar = featureVarImg(x, y);
// Filter using pixels within search range
for(int dy = -searchRadius; dy <= searchRadius; dy++)
for(int dx = -searchRadius; dx <= searchRadius; dx++) {
int xx = x + dx;
int yy = y + dy;
if(xx < 0 || yy < 0 || xx >= width || yy >= height)
continue;
Color diffSqSum(0.f, 0.f, 0.f);
// Calculate block distance
for(int by = -patchRadius; by <= patchRadius; by++)
for(int bx = -patchRadius; bx <= patchRadius; bx++) {
int xbx = x + bx;
int yby = y + by;
int xxbx = xx + bx;
int yyby = yy + by;
if( xbx < 0 || xbx >= width ||
yby < 0 || yby >= height ||
xxbx < 0 || xxbx >= width ||
yyby < 0 || yyby >= height)
continue;
Color diff = rImg(xbx, yby) - rImg(xxbx, yyby);
diffSqSum += (diff*diff);
}
diffSqSum *= invPatchSize;
float dist = Avg(diffSqSum);
Feature fDiff = feature - featureImg(xx, yy);
Feature fVarSum = featureVar + featureVarImg(xx, yy);
Feature fDist = (fDiff*fDiff)/fVarSum.Max(c_VarMax);
// For each parameter, calculate information necessary for
// filtering and SURE estimation
for(size_t p = 0; p < scaleR.size(); p++) {
if(scaleR[p] == 0.f) {
continue;
}
float weight = fmath::exp(dist*scaleR[p] +
Sum(fDist*scaleF));
sum[p] += weight * img(xx, yy);
rSum[p] += weight * rImg(xx, yy);
rSumSq[p] += weight * rImg(xx, yy) * rImg(xx, yy);
wSum[p] += weight;
if(dy >= -patchRadius && dy <= patchRadius &&
dx >= -patchRadius && dx <= patchRadius) {
wArray[p][(dy+patchRadius)*patchWidth+(dx+patchRadius)] =
weight;
}
}
}
for(size_t p = 0; p < scaleR.size(); p++) {
if(scaleR[p] == 0.f) {
fltArray[p](x, y) = img(x, y);
mseArray[p](x, y) = Avg(2.f*varImg(x, y));
continue;
}
float invWSum = 1.f/wSum[p];
Color xl = sum[p] * invWSum;
Color rxl = rSum[p] * invWSum;
Color rxlSq = rSumSq[p] * invWSum;
Color ryl = rImg(x, y);
Color dxdy = (-2.f*scaleR[p])*(rxlSq - rxl*rxl)*invPatchSize + Color(invWSum);
Color tmp;
for(int by = -patchRadius; by <= patchRadius; by++)
for(int bx = -patchRadius; bx <= patchRadius; bx++) {
int xbpx = x+bx;
int ybpy = y+by;
int xbmx = x-bx;
int ybmy = y-by;
if( xbpx < 0 || xbpx >= width ||
ybpy < 0 || ybpy >= height ||
xbmx < 0 || xbmx >= width ||
ybmy < 0 || ybmy >= height)
continue;
Color rylpb = rImg(xbpx, ybpy);
Color rylmb = rImg(xbmx, ybmy);
float w = wArray[p][(-by+patchRadius)*patchWidth+(-bx+patchRadius)];
tmp += w*(ryl - rylpb)*(rxl - rylmb);
}
tmp *= (-2.f*scaleR[p])*invPatchSize*invWSum;
dxdy += tmp;
Color mse = (rxl-ryl)*(rxl-ryl) + 2.f*varImg(x, y)*dxdy - varImg(x, y);
Color pri = (mse + varImg(x, y));
fltArray[p](x, y) = xl;
mseArray[p](x, y) = Avg(mse);
priArray[p](x, y) = Avg(pri) / (xl.Y()*xl.Y() + 1e-2f);
}
}
}
}
