void HighPassFilter::applyLowPass(const Image2DPtr &image)
{
// Guassian convolution can be separated in two 1D convolution
// because of properties of the 2D Gaussian function.
Image2DPtr temp = Image2D::CreateZeroImagePtr(image->Width(), image->Height());
size_t hKernelMid = _hWindowSize/2;
for(size_t i=0; i<_hWindowSize; ++i) {
const num_t kernelValue = _hKernel[i];
const size_t
xStart = (i >= hKernelMid) ? 0 : (hKernelMid-i),
xEnd = (i <= hKernelMid) ? image->Width() : image->Width()-i+hKernelMid;
for(unsigned y=0;y<image->Height();++y) {
for(unsigned x=xStart;x<xEnd;++x)
temp->AddValue(x, y, image->Value(x+i-hKernelMid, y)*kernelValue);
}
}
image->SetAll(0.0);
size_t vKernelMid = _vWindowSize/2;
for(size_t i=0; i<_vWindowSize; ++i) {
const num_t kernelValue = _vKernel[i];
const size_t
yStart = (i >= vKernelMid) ? 0 : (vKernelMid-i),
yEnd = (i <= vKernelMid) ? image->Height() : image->Height()-i+vKernelMid;
for(unsigned y=yStart;y<yEnd;++y) {
for(unsigned x=0;x<image->Width();++x)
image->AddValue(x, y, temp->Value(x, y+i-vKernelMid)*kernelValue);
}
}
}
void TimeConvolutionAction::PerformFFTSincOperation(ArtifactSet &artifacts, Image2DPtr real, Image2DPtr imag) const
{
fftw_complex
*fftIn = (fftw_complex*) fftw_malloc(sizeof(fftw_complex) * real->Width()),
*fftOut = (fftw_complex*) fftw_malloc(sizeof(fftw_complex) * real->Width());
// FFTW plan routines are not thread safe, so lock.
boost::mutex::scoped_lock lock(artifacts.IOMutex());
fftw_plan
fftPlanForward = fftw_plan_dft_1d(real->Width(), fftIn, fftOut, FFTW_FORWARD, FFTW_MEASURE),
fftPlanBackward = fftw_plan_dft_1d(real->Width(), fftIn, fftOut, FFTW_BACKWARD, FFTW_MEASURE);
lock.unlock();
const size_t width = real->Width();
const BandInfo band = artifacts.MetaData()->Band();
for(unsigned y=0;y<real->Height();++y)
{
const numl_t sincScale = ActualSincScaleInSamples(artifacts, band.channels[y].frequencyHz);
const numl_t limitFrequency = (numl_t) width / sincScale;
if(y == real->Height()/2)
{
AOLogger::Debug << "Horizontal sinc scale: " << sincScale << " (filter scale: " << Angle::ToString(ActualSincScaleAsRaDecDist(artifacts, band.channels[y].frequencyHz)) << ")\n";
}
if(sincScale > 1.0)
{
for(unsigned x=0;x<width;++x)
{
fftIn[x][0] = real->Value(x, y);
fftIn[x][1] = imag->Value(x, y);
}
fftw_execute_dft(fftPlanForward, fftIn, fftOut);
size_t filterIndexSize = (limitFrequency > 1.0) ? (size_t) ceil(limitFrequency/2.0) : 1;
// Remove the high frequencies [filterIndexSize : n-filterIndexSize]
for(size_t f=filterIndexSize;f<width - filterIndexSize;++f)
{
fftOut[f][0] = 0.0;
fftOut[f][1] = 0.0;
}
fftw_execute_dft(fftPlanBackward, fftOut, fftIn);
const double n = width;
for(unsigned x=0;x<width;++x)
{
real->SetValue(x, y, fftIn[x][0] / n);
imag->SetValue(x, y, fftIn[x][1] / n);
}
}
}
fftw_free(fftIn);
fftw_free(fftOut);
}
void HighPassFilter::elementWiseDivide(const Image2DPtr &leftHand, const Image2DCPtr &rightHand)
{
for(unsigned y=0;y<leftHand->Height();++y) {
for(unsigned x=0;x<leftHand->Width();++x) {
if(rightHand->Value(x, y) == 0.0)
leftHand->SetValue(x, y, 0.0);
else
leftHand->SetValue(x, y, leftHand->Value(x, y) / rightHand->Value(x, y));
}
}
}
num_t SpatialCompositionAction::eigenvalue(Image2DCPtr real, Image2DCPtr imaginary) const
{
try {
Image2DPtr
r = Image2D::CreateCopy(real),
i = Image2D::CreateCopy(imaginary);
for(size_t y=0;y<r->Height();++y)
{
for(size_t x=0;x<r->Width();++x)
{
if(!std::isfinite(r->Value(x,y))) r->SetValue(x, y, 0.0);
if(!std::isfinite(i->Value(x,y))) i->SetValue(x, y, 0.0);
}
}
if(r->ContainsOnlyZeros() && i->ContainsOnlyZeros()) return 0.0;
return Eigenvalue::Compute(r, i);
} catch(std::exception &e)
{
return std::numeric_limits<num_t>::quiet_NaN();
}
}
void HighPassFilter::elementWiseDivideSSE(const Image2DPtr &leftHand, const Image2DCPtr &rightHand)
{
const __m128 zero4 = _mm_set_ps(0.0, 0.0, 0.0, 0.0);
for(unsigned y=0;y<leftHand->Height();++y) {
float *leftHandPtr = leftHand->ValuePtr(0, y);
const float *rightHandPtr = rightHand->ValuePtr(0, y);
float *end = leftHandPtr + leftHand->Width();
while(leftHandPtr < end)
{
__m128
l = _mm_load_ps(leftHandPtr),
r = _mm_load_ps(rightHandPtr);
__m128 conditionMask = _mm_cmpeq_ps(r, zero4);
_mm_store_ps(leftHandPtr, _mm_or_ps(
_mm_and_ps(conditionMask, zero4),
_mm_andnot_ps(conditionMask, _mm_div_ps(l, r))
));
leftHandPtr += 4;
rightHandPtr += 4;
}
}
}
void HighPassFilter::applyLowPassSSE(const Image2DPtr &image)
{
Image2DPtr temp = Image2D::CreateZeroImagePtr(image->Width(), image->Height());
unsigned hKernelMid = _hWindowSize/2;
for(unsigned i=0; i<_hWindowSize; ++i) {
const num_t k = _hKernel[i];
const __m128 k4 = _mm_set_ps(k, k, k, k);
unsigned
/* xStart is the first column to start writing to. Note that it might be larger
* than the width. */
xStart = (i >= hKernelMid) ? 0 : (hKernelMid-i),
xEnd = (i <= hKernelMid) ? image->Width() : (image->Width()+hKernelMid > i ? (image->Width()-i+hKernelMid) : 0);
for(unsigned y=0;y<image->Height();++y) {
float *tempPtr = temp->ValuePtr(xStart, y);
const float *imagePtr = image->ValuePtr(xStart+i-hKernelMid, y);
unsigned x = xStart;
for(;x+4<xEnd;x+=4) {
const __m128
imageVal = _mm_loadu_ps(imagePtr),
tempVal = _mm_loadu_ps(tempPtr);
// *tempPtr += k * (*imagePtr);
_mm_storeu_ps(tempPtr, _mm_add_ps(tempVal, _mm_mul_ps(imageVal, k4)));
tempPtr += 4;
imagePtr += 4;
}
for(;x<xEnd;++x) {
*tempPtr += k * (*imagePtr);
++tempPtr;
++imagePtr;
}
}
}
image->SetAll(0.0);
unsigned vKernelMid = _vWindowSize/2;
for(unsigned i=0; i<_vWindowSize; ++i) {
const num_t k = _vKernel[i];
const __m128 k4 = _mm_set_ps(k, k, k, k);
const unsigned
yStart = (i >= vKernelMid) ? 0 : (vKernelMid-i),
yEnd = (i <= vKernelMid) ? image->Height() : ((image->Height()+vKernelMid>i) ? (image->Height()-i+vKernelMid) : 0);
for(unsigned y=yStart;y<yEnd;++y) {
const float *tempPtr = temp->ValuePtr(0, y+i-vKernelMid);
float *imagePtr = image->ValuePtr(0, y);
unsigned x=0;
for(;x+4<image->Width();x += 4) {
const __m128
imageVal = _mm_load_ps(imagePtr),
tempVal = _mm_load_ps(tempPtr);
// *imagePtr += k * (*tempPtr);
_mm_store_ps(imagePtr, _mm_add_ps(imageVal, _mm_mul_ps(tempVal, k4)));
tempPtr += 4;
imagePtr += 4;
}
for(;x<image->Width();++x) {
*imagePtr += k * (*tempPtr);
++tempPtr;
++imagePtr;
}
}
}
}
