Speedup und reduced memory usage for Noise Reduction, Issue 2557 #132
This commit is contained in:
Ingo
@@ -31,9 +31,6 @@ namespace rtengine {
|
||||
class wavelet_level
|
||||
{
|
||||
|
||||
// size of padded border
|
||||
size_t m_pad;
|
||||
|
||||
// level of decomposition
|
||||
int lvl;
|
||||
|
||||
@@ -49,19 +46,27 @@ namespace rtengine {
|
||||
|
||||
// load a row/column of input data, possibly with padding
|
||||
|
||||
void AnalysisFilterHaarVertical (T * srcbuffer, T * dstLo, T * dstHi, int pitch, int srclen, int row);
|
||||
void AnalysisFilterHaarHorizontal (T * srcbuffer, T * dstLo, T * dstHi, int srclen, int row);
|
||||
void SynthesisFilterHaarHorizontal (T * srcLo, T * srcHi, T * dst, int dstlen);
|
||||
void SynthesisFilterHaarVertical (T * srcLo, T * srcHi, T * dst, int pitch, int dstlen);
|
||||
void AnalysisFilterHaarVertical (const T * const srcbuffer, T * dstLo, T * dstHi, const int width, const int height, const int row);
|
||||
void AnalysisFilterHaarHorizontal (const T * const srcbuffer, T * dstLo, T * dstHi, const int width, const int row);
|
||||
void SynthesisFilterHaarHorizontal (const T * const srcLo, const T * const srcHi, T * dst, const int width, const int height);
|
||||
void SynthesisFilterHaarVertical (const T * const srcLo, const T * const srcHi, T * dst, const int width, const int height);
|
||||
|
||||
void AnalysisFilterSubsampHorizontal (T * srcbuffer, T * dstLo, T * dstHi, float *filterLo, float *filterHi,
|
||||
int taps, int offset, int pitch, int srclen, int m_w2, int row);
|
||||
const int taps, const int offset, const int srcwidth, const int dstwidth, const int row);
|
||||
#ifdef __SSE2__
|
||||
void AnalysisFilterSubsampVertical (T * srcbuffer, T * dstLo, T * dstHi, float (*filterLo)[4], float (*filterHi)[4],
|
||||
const int taps, const int offset, const int width, const int height, const int row);
|
||||
#else
|
||||
void AnalysisFilterSubsampVertical (T * srcbuffer, T * dstLo, T * dstHi, float *filterLo, float *filterHi,
|
||||
int taps, int offset, int pitch, int srclen, int row);
|
||||
int const taps, const int offset, const int width, const int height, const int row);
|
||||
#endif
|
||||
void SynthesisFilterSubsampHorizontal (T * srcLo, T * srcHi, T * dst,
|
||||
float *filterLo, float *filterHi, int taps, int offset, int dstlen);
|
||||
void SynthesisFilterSubsampVertical (T * srcLo, T * srcHi, T * dst, float *filterLo, float *filterHi, int taps, int offset, int pitch, int dstlen);
|
||||
|
||||
float *filterLo, float *filterHi, const int taps, const int offset, const int scrwidth, const int dstwidth, const int height);
|
||||
#ifdef __SSE2__
|
||||
void SynthesisFilterSubsampVertical (T * srcLo, T * srcHi, T * dst, float (*filterLo)[4], float (*filterHi)[4], const int taps, const int offset, const int width, const int srcheight, const int dstheight);
|
||||
#else
|
||||
void SynthesisFilterSubsampVertical (T * srcLo, T * srcHi, T * dst, float *filterLo, float *filterHi, const int taps, const int offset, const int width, const int srcheight, const int dstheight);
|
||||
#endif
|
||||
public:
|
||||
|
||||
T ** wavcoeffs;
|
||||
@@ -72,18 +77,20 @@ namespace rtengine {
|
||||
size_t m_w2, m_h2;
|
||||
|
||||
template<typename E>
|
||||
wavelet_level(E * src, E * dst, int level, int subsamp, int padding, size_t w, size_t h, float *filterV, float *filterH, int len, int offset)
|
||||
: m_w(w), m_h(h), m_w2(w), m_h2(h), m_pad(padding), wavcoeffs(NULL), lvl(level), skip(1<<level), subsamp_out((subsamp>>level)&1)
|
||||
wavelet_level(E * src, E * dst, int level, int subsamp, size_t w, size_t h, float *filterV, float *filterH, int len, int offset, int skipcrop)
|
||||
: lvl(level), subsamp_out((subsamp>>level)&1), skip(1<<level), wavcoeffs(NULL), m_w(w), m_h(h), m_w2(w), m_h2(h)
|
||||
{
|
||||
if (subsamp) {
|
||||
skip = 1;
|
||||
for (int n=0; n<level; n++) {
|
||||
skip *= 2-((subsamp>>n)&1);
|
||||
}
|
||||
}
|
||||
skip /= skipcrop;
|
||||
if(skip < 1) skip=1;
|
||||
|
||||
}
|
||||
m_w2 = (subsamp_out ? ((w+1+2*skip*padding)/2) : (w+2*skip*padding));
|
||||
m_h2 = (subsamp_out ? ((h+1+2*skip*padding)/2) : (h+2*skip*padding));
|
||||
m_pad= skip*padding;
|
||||
m_w2 = (subsamp_out ? (w+1)/2 : w);
|
||||
m_h2 = (subsamp_out ? (h+1)/2 : h);
|
||||
|
||||
wavcoeffs = create((m_w2)*(m_h2));
|
||||
decompose_level(src, dst, filterV, filterH, len, offset);
|
||||
@@ -115,11 +122,6 @@ namespace rtengine {
|
||||
return m_h2;
|
||||
}
|
||||
|
||||
size_t padding() const
|
||||
{
|
||||
return m_pad/skip;
|
||||
}
|
||||
|
||||
size_t stride() const
|
||||
{
|
||||
return skip;
|
||||
@@ -151,33 +153,33 @@ namespace rtengine {
|
||||
}
|
||||
|
||||
template<typename T>
|
||||
void wavelet_level<T>::AnalysisFilterHaarHorizontal (T * RESTRICT srcbuffer, T * RESTRICT dstLo, T * RESTRICT dstHi, int srclen, int row) {
|
||||
void wavelet_level<T>::AnalysisFilterHaarHorizontal (const T * const RESTRICT srcbuffer, T * RESTRICT dstLo, T * RESTRICT dstHi, const int width, const int row) {
|
||||
/* Basic convolution code
|
||||
* Applies a Haar filter
|
||||
*/
|
||||
for(int i = 0; i < (srclen - skip); i++) {
|
||||
dstLo[row*srclen+i] = (srcbuffer[i] + srcbuffer[i+skip]);
|
||||
dstHi[row*srclen+i] = (srcbuffer[i] - srcbuffer[i+skip]);
|
||||
for(int i = 0; i < (width - skip); i++) {
|
||||
dstLo[row*width+i] = (srcbuffer[i] + srcbuffer[i+skip]);
|
||||
dstHi[row*width+i] = (srcbuffer[i] - srcbuffer[i+skip]);
|
||||
}
|
||||
for(size_t i = max(srclen-skip,skip); i < (srclen); i++) {
|
||||
dstLo[row*srclen+i] = (srcbuffer[i] + srcbuffer[i-skip]);
|
||||
dstHi[row*srclen+i] = (srcbuffer[i] - srcbuffer[i-skip]);
|
||||
for(size_t i = max(width-skip,skip); i < (width); i++) {
|
||||
dstLo[row*width+i] = (srcbuffer[i] + srcbuffer[i-skip]);
|
||||
dstHi[row*width+i] = (srcbuffer[i] - srcbuffer[i-skip]);
|
||||
}
|
||||
}
|
||||
|
||||
template<typename T> void wavelet_level<T>::AnalysisFilterHaarVertical (T * RESTRICT srcbuffer, T * RESTRICT dstLo, T * RESTRICT dstHi, int pitch, int srclen, int row) {
|
||||
template<typename T> void wavelet_level<T>::AnalysisFilterHaarVertical (const T * const RESTRICT srcbuffer, T * RESTRICT dstLo, T * RESTRICT dstHi, const int width, const int height, const int row) {
|
||||
/* Basic convolution code
|
||||
* Applies a Haar filter
|
||||
*/
|
||||
if(row < (srclen - skip)) {
|
||||
for(int j=0;j<pitch;j++) {
|
||||
dstLo[j] = 0.25f*(srcbuffer[row*pitch+j] + srcbuffer[(row+skip)*pitch+j]);
|
||||
dstHi[j] = 0.25f*(srcbuffer[row*pitch+j] - srcbuffer[(row+skip)*pitch+j]);
|
||||
if(row < (height - skip)) {
|
||||
for(int j=0;j<width;j++) {
|
||||
dstLo[j] = 0.25f*(srcbuffer[row*width+j] + srcbuffer[(row+skip)*width+j]);
|
||||
dstHi[j] = 0.25f*(srcbuffer[row*width+j] - srcbuffer[(row+skip)*width+j]);
|
||||
}
|
||||
} else if(row>=max(srclen-skip,skip)) {
|
||||
for(int j=0;j<pitch;j++) {
|
||||
dstLo[j] = 0.25f*(srcbuffer[row*pitch+j] + srcbuffer[(row-skip)*pitch+j]);
|
||||
dstHi[j] = 0.25f*(srcbuffer[row*pitch+j] - srcbuffer[(row-skip)*pitch+j]);
|
||||
} else if(row>=max(height-skip,skip)) {
|
||||
for(int j=0;j<width;j++) {
|
||||
dstLo[j] = 0.25f*(srcbuffer[row*width+j] + srcbuffer[(row-skip)*width+j]);
|
||||
dstHi[j] = 0.25f*(srcbuffer[row*width+j] - srcbuffer[(row-skip)*width+j]);
|
||||
}
|
||||
}
|
||||
}
|
||||
@@ -185,43 +187,43 @@ namespace rtengine {
|
||||
// %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|
||||
// %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|
||||
|
||||
template<typename T> void wavelet_level<T>::SynthesisFilterHaarHorizontal (T * RESTRICT srcLo, T * RESTRICT srcHi, T * RESTRICT dst, int dstlen) {
|
||||
template<typename T> void wavelet_level<T>::SynthesisFilterHaarHorizontal (const T * const RESTRICT srcLo, const T * const RESTRICT srcHi, T * RESTRICT dst, const int width, const int height) {
|
||||
|
||||
/* Basic convolution code
|
||||
* Applies a Haar filter
|
||||
*
|
||||
*/
|
||||
|
||||
for (int k=0; k<m_h2; k++) {
|
||||
for(size_t i = (m_pad); i < (m_pad+skip); i++) {
|
||||
dst[k*m_w+(i-m_pad)] = (srcLo[k*m_w2+i] + srcHi[k*m_w2+i]);
|
||||
for (int k=0; k<height; k++) {
|
||||
for(size_t i = 0; i < skip; i++) {
|
||||
dst[k*width+i] = (srcLo[k*width+i] + srcHi[k*width+i]);
|
||||
}
|
||||
for(size_t i = m_pad+skip; i < (dstlen+m_pad); i++) {
|
||||
dst[k*m_w+(i-m_pad)] = 0.5f*(srcLo[k*m_w2+i] + srcHi[k*m_w2+i] + srcLo[k*m_w2+i-skip] - srcHi[k*m_w2+i-skip]);
|
||||
for(size_t i = skip; i < width; i++) {
|
||||
dst[k*width+i] = 0.5f*(srcLo[k*width+i] + srcHi[k*width+i] + srcLo[k*width+i-skip] - srcHi[k*width+i-skip]);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
template<typename T> void wavelet_level<T>::SynthesisFilterHaarVertical (T * RESTRICT srcLo, T * RESTRICT srcHi, T * RESTRICT dst, int pitch, int dstlen) {
|
||||
template<typename T> void wavelet_level<T>::SynthesisFilterHaarVertical (const T * const RESTRICT srcLo, const T * const RESTRICT srcHi, T * RESTRICT dst, const int width, const int height) {
|
||||
|
||||
/* Basic convolution code
|
||||
* Applies a Haar filter
|
||||
*
|
||||
*/
|
||||
|
||||
for(size_t i = (m_pad); i < (m_pad+skip); i++) {
|
||||
for(int j=0;j<pitch;j++)
|
||||
dst[pitch*(i-m_pad)+j] = (srcLo[i*pitch+j] + srcHi[i*pitch+j]);
|
||||
for(size_t i = 0; i < skip; i++) {
|
||||
for(int j=0;j<width;j++)
|
||||
dst[width*i+j] = (srcLo[i*width+j] + srcHi[i*width+j]);
|
||||
}
|
||||
for(size_t i = m_pad+skip; i < (dstlen+m_pad); i++) {
|
||||
for(int j=0;j<pitch;j++)
|
||||
dst[pitch*(i-m_pad)+j] = 0.5f*(srcLo[i*pitch+j] + srcHi[i*pitch+j] + srcLo[(i-skip)*pitch+j] - srcHi[(i-skip)*pitch+j]);
|
||||
for(size_t i = skip; i < height; i++) {
|
||||
for(int j=0;j<width;j++)
|
||||
dst[width*i+j] = 0.5f*(srcLo[i*width+j] + srcHi[i*width+j] + srcLo[(i-skip)*width+j] - srcHi[(i-skip)*width+j]);
|
||||
}
|
||||
}
|
||||
|
||||
template<typename T>
|
||||
void wavelet_level<T>::AnalysisFilterSubsampHorizontal (T * RESTRICT srcbuffer, T * RESTRICT dstLo, T * RESTRICT dstHi, float * RESTRICT filterLo, float *filterHi,
|
||||
int taps, int offset, int pitch, int srclen, int m_w2, int row) {
|
||||
void wavelet_level<T>::AnalysisFilterSubsampHorizontal (T * RESTRICT srcbuffer, T * RESTRICT dstLo, T * RESTRICT dstHi, float * RESTRICT filterLo, float *RESTRICT filterHi,
|
||||
const int taps, const int offset, const int srcwidth, const int dstwidth, const int row) {
|
||||
/* Basic convolution code
|
||||
* Applies an FIR filter 'filter' with filter length 'taps',
|
||||
* aligning the 'offset' element of the filter with
|
||||
@@ -229,62 +231,125 @@ namespace rtengine {
|
||||
* Output is subsampled by two
|
||||
*/
|
||||
// calculate coefficients
|
||||
for(int i = 0; i < srclen; i+=2) {
|
||||
float lo = 0.f, hi = 0.f;
|
||||
if (LIKELY(i>skip*taps && i<srclen-skip*taps)) {//bulk
|
||||
for (int j=0, l=-skip*offset; j<taps; j++, l+=skip) {
|
||||
float src = srcbuffer[i-l];
|
||||
lo += filterLo[j] * src;//lopass channel
|
||||
hi += filterHi[j] * src;//hipass channel
|
||||
}
|
||||
} else {
|
||||
for (int j=0; j<taps; j++) {
|
||||
int arg = max(0,min(i+skip*(offset-j),srclen-1));//clamped BC's
|
||||
lo += filterLo[j] * srcbuffer[arg];//lopass channel
|
||||
hi += filterHi[j] * srcbuffer[arg];//hipass channel
|
||||
}
|
||||
}
|
||||
dstLo[row*m_w2+((i/2))] = lo;
|
||||
dstHi[row*m_w2+((i/2))] = hi;
|
||||
}
|
||||
}
|
||||
|
||||
template<typename T> void wavelet_level<T>::AnalysisFilterSubsampVertical (T * RESTRICT srcbuffer, T * RESTRICT dstLo, T * RESTRICT dstHi, float * RESTRICT filterLo, float * RESTRICT filterHi,
|
||||
int taps, int offset, int pitch, int srclen, int row) {
|
||||
|
||||
/* Basic convolution code
|
||||
* Applies an FIR filter 'filter' with filter length 'taps',
|
||||
* aligning the 'offset' element of the filter with
|
||||
* the input pixel, and skipping 'skip' pixels between taps
|
||||
* Output is subsampled by two
|
||||
*/
|
||||
|
||||
// calculate coefficients
|
||||
if (LIKELY(row>skip*taps && row<srclen-skip*taps)) {//bulk
|
||||
for (int k=0; k<pitch; k++) {
|
||||
float lo = 0.f, hi = 0.f;
|
||||
for (int j=0, l=-skip*offset; j<taps; j++, l+=skip) {
|
||||
lo += filterLo[j] * srcbuffer[(row-l)*pitch+k];//lopass channel
|
||||
hi += filterHi[j] * srcbuffer[(row-l)*pitch+k];//hipass channel
|
||||
}
|
||||
dstLo[k] = lo;
|
||||
dstHi[k] = hi;
|
||||
}
|
||||
} else {//boundary
|
||||
for (int k=0; k<pitch; k++) {
|
||||
float lo = 0.f, hi = 0.f;
|
||||
for (int j=0; j<taps; j++) {
|
||||
int arg = max(0,min(row+skip*(offset-j),srclen-1))*pitch+k;//clamped BC's
|
||||
lo += filterLo[j] * srcbuffer[arg];//lopass channel
|
||||
hi += filterHi[j] * srcbuffer[arg];//hipass channel
|
||||
}
|
||||
dstLo[k] = lo;
|
||||
dstHi[k] = hi;
|
||||
for(int i = 0; i < srcwidth; i+=2) {
|
||||
float lo = 0.f, hi = 0.f;
|
||||
if (LIKELY(i>skip*taps && i<srcwidth-skip*taps)) {//bulk
|
||||
for (int j=0, l=-skip*offset; j<taps; j++, l+=skip) {
|
||||
float src = srcbuffer[i-l];
|
||||
lo += filterLo[j] * src;//lopass channel
|
||||
hi += filterHi[j] * src;//hipass channel
|
||||
}
|
||||
} else {
|
||||
for (int j=0; j<taps; j++) {
|
||||
int arg = max(0,min(i+skip*(offset-j),srcwidth-1));//clamped BC's
|
||||
lo += filterLo[j] * srcbuffer[arg];//lopass channel
|
||||
hi += filterHi[j] * srcbuffer[arg];//hipass channel
|
||||
}
|
||||
}
|
||||
dstLo[row*dstwidth+((i/2))] = lo;
|
||||
dstHi[row*dstwidth+((i/2))] = hi;
|
||||
}
|
||||
}
|
||||
|
||||
template<typename T> void wavelet_level<T>::SynthesisFilterSubsampHorizontal (T * RESTRICT srcLo, T * RESTRICT srcHi, T * RESTRICT dst, float * RESTRICT filterLo, float * RESTRICT filterHi, int taps, int offset, int dstlen) {
|
||||
#ifdef __SSE2__
|
||||
template<typename T> void wavelet_level<T>::AnalysisFilterSubsampVertical (T * RESTRICT srcbuffer, T * RESTRICT dstLo, T * RESTRICT dstHi, float (* RESTRICT filterLo)[4], float (* RESTRICT filterHi)[4],
|
||||
const int taps, const int offset, const int width, const int height, const int row) {
|
||||
|
||||
/* Basic convolution code
|
||||
* Applies an FIR filter 'filter' with filter length 'taps',
|
||||
* aligning the 'offset' element of the filter with
|
||||
* the input pixel, and skipping 'skip' pixels between taps
|
||||
* Output is subsampled by two
|
||||
*/
|
||||
|
||||
// calculate coefficients
|
||||
if (LIKELY(row>skip*taps && row<height-skip*taps)) {//bulk
|
||||
int k;
|
||||
for (k=0; k<width-3; k+=4) {
|
||||
__m128 lov = _mm_setzero_ps();
|
||||
__m128 hiv = _mm_setzero_ps();
|
||||
for (int j=0, l=-skip*offset; j<taps; j++, l+=skip) {
|
||||
__m128 srcv = LVFU(srcbuffer[(row-l)*width+k]);
|
||||
lov += LVF(filterLo[j][0]) * srcv;//lopass channel
|
||||
hiv += LVF(filterHi[j][0]) * srcv;//hipass channel
|
||||
}
|
||||
STVF(dstLo[k], lov);
|
||||
STVF(dstHi[k], hiv);
|
||||
}
|
||||
for (; k<width; k++) {
|
||||
float lo = 0.f, hi = 0.f;
|
||||
for (int j=0, l=-skip*offset; j<taps; j++, l+=skip) {
|
||||
lo += filterLo[j][0] * srcbuffer[(row-l)*width+k];//lopass channel
|
||||
hi += filterHi[j][0] * srcbuffer[(row-l)*width+k];//hipass channel
|
||||
}
|
||||
dstLo[k] = lo;
|
||||
dstHi[k] = hi;
|
||||
}
|
||||
} else {//boundary
|
||||
int k;
|
||||
for (k=0; k<width-3; k+=4) {
|
||||
__m128 lov = _mm_setzero_ps();
|
||||
__m128 hiv = _mm_setzero_ps();
|
||||
for (int j=0; j<taps; j++) {
|
||||
int arg = max(0,min(row+skip*(offset-j),height-1))*width+k;//clamped BC's
|
||||
__m128 srcv = LVFU(srcbuffer[arg]);
|
||||
lov += LVF(filterLo[j][0]) * srcv;//lopass channel
|
||||
hiv += LVF(filterHi[j][0]) * srcv;//hipass channel
|
||||
}
|
||||
STVF(dstLo[k], lov);
|
||||
STVF(dstHi[k], hiv);
|
||||
}
|
||||
for (; k<width; k++) {
|
||||
float lo = 0.f, hi = 0.f;
|
||||
for (int j=0; j<taps; j++) {
|
||||
int arg = max(0,min(row+skip*(offset-j),height-1))*width+k;//clamped BC's
|
||||
lo += filterLo[j][0] * srcbuffer[arg];//lopass channel
|
||||
hi += filterHi[j][0] * srcbuffer[arg];//hipass channel
|
||||
}
|
||||
dstLo[k] = lo;
|
||||
dstHi[k] = hi;
|
||||
}
|
||||
}
|
||||
}
|
||||
#else
|
||||
template<typename T> void wavelet_level<T>::AnalysisFilterSubsampVertical (T * RESTRICT srcbuffer, T * RESTRICT dstLo, T * RESTRICT dstHi, float * RESTRICT filterLo, float * RESTRICT filterHi,
|
||||
const int taps, const int offset, const int width, const int height, const int row) {
|
||||
|
||||
/* Basic convolution code
|
||||
* Applies an FIR filter 'filter' with filter length 'taps',
|
||||
* aligning the 'offset' element of the filter with
|
||||
* the input pixel, and skipping 'skip' pixels between taps
|
||||
* Output is subsampled by two
|
||||
*/
|
||||
|
||||
// calculate coefficients
|
||||
if (LIKELY(row>skip*taps && row<height-skip*taps)) {//bulk
|
||||
for (int k=0; k<width; k++) {
|
||||
float lo = 0.f, hi = 0.f;
|
||||
for (int j=0, l=-skip*offset; j<taps; j++, l+=skip) {
|
||||
lo += filterLo[j] * srcbuffer[(row-l)*width+k];//lopass channel
|
||||
hi += filterHi[j] * srcbuffer[(row-l)*width+k];//hipass channel
|
||||
}
|
||||
dstLo[k] = lo;
|
||||
dstHi[k] = hi;
|
||||
}
|
||||
} else {//boundary
|
||||
for (int k=0; k<width; k++) {
|
||||
float lo = 0.f, hi = 0.f;
|
||||
for (int j=0; j<taps; j++) {
|
||||
int arg = max(0,min(row+skip*(offset-j),height-1))*width+k;//clamped BC's
|
||||
lo += filterLo[j] * srcbuffer[arg];//lopass channel
|
||||
hi += filterHi[j] * srcbuffer[arg];//hipass channel
|
||||
}
|
||||
dstLo[k] = lo;
|
||||
dstHi[k] = hi;
|
||||
}
|
||||
}
|
||||
}
|
||||
#endif
|
||||
|
||||
|
||||
template<typename T> void wavelet_level<T>::SynthesisFilterSubsampHorizontal (T * RESTRICT srcLo, T * RESTRICT srcHi, T * RESTRICT dst, float * RESTRICT filterLo, float * RESTRICT filterHi, const int taps, const int offset, const int srcwidth, const int dstwidth, const int height) {
|
||||
|
||||
/* Basic convolution code
|
||||
* Applies an FIR filter 'filter' with filter length 'taps',
|
||||
@@ -294,78 +359,157 @@ namespace rtengine {
|
||||
*/
|
||||
|
||||
// calculate coefficients
|
||||
int srclen = (dstlen==m_w ? m_w2 : m_h2);//length of row/col in src (coarser level)
|
||||
int shift = skip*(taps-offset-1);//align filter with data
|
||||
|
||||
for (int k=0; k<m_h2; k++) {
|
||||
for(size_t i = m_pad; i < (dstlen+m_pad); i++) {
|
||||
for (int k=0; k<height; k++) {
|
||||
int i;
|
||||
for(i=0; i<=min(skip*taps,dstwidth); i++) {
|
||||
float tot=0.f;
|
||||
//TODO: this is correct only if skip=1; otherwise, want to work with cosets of length 'skip'
|
||||
int i_src = (i+shift)/2;
|
||||
int begin = (i+shift)%2;
|
||||
if (LIKELY(i>skip*taps && i<(srclen-skip*taps))) {//bulk
|
||||
for (int j=begin, l=0; j<taps; j+=2, l+=skip) {
|
||||
tot += ((filterLo[j] * srcLo[k*m_w2+i_src-l] + filterHi[j] * srcHi[k*m_w2+i_src-l]));
|
||||
}
|
||||
} else {//boundary
|
||||
for (int j=begin, l=0; j<taps; j+=2, l+=skip) {
|
||||
int arg = max(0,min((i_src-l),srclen-1));//clamped BC's
|
||||
tot += ((filterLo[j] * srcLo[k*m_w2+arg] + filterHi[j] * srcHi[k*m_w2+arg]));
|
||||
}
|
||||
for (int j=begin, l=0; j<taps; j+=2, l+=skip) {
|
||||
int arg = max(0,min((i_src-l),srcwidth-1));//clamped BC's
|
||||
tot += ((filterLo[j] * srcLo[k*srcwidth+arg] + filterHi[j] * srcHi[k*srcwidth+arg]));
|
||||
}
|
||||
dst[k*m_w+(i-m_pad)] = tot;
|
||||
dst[k*dstwidth+i] = tot;
|
||||
}
|
||||
for(; i<min(dstwidth-skip*taps,dstwidth); i++) {
|
||||
float tot=0.f;
|
||||
//TODO: this is correct only if skip=1; otherwise, want to work with cosets of length 'skip'
|
||||
int i_src = (i+shift)/2;
|
||||
int begin = (i+shift)%2;
|
||||
for (int j=begin, l=0; j<taps; j+=2, l+=skip) {
|
||||
tot += ((filterLo[j] * srcLo[k*srcwidth+i_src-l] + filterHi[j] * srcHi[k*srcwidth+i_src-l]));
|
||||
}
|
||||
dst[k*dstwidth+i] = tot;
|
||||
}
|
||||
for(; i < dstwidth; i++) {
|
||||
float tot=0.f;
|
||||
//TODO: this is correct only if skip=1; otherwise, want to work with cosets of length 'skip'
|
||||
int i_src = (i+shift)/2;
|
||||
int begin = (i+shift)%2;
|
||||
for (int j=begin, l=0; j<taps; j+=2, l+=skip) {
|
||||
int arg = max(0,min((i_src-l),srcwidth-1));//clamped BC's
|
||||
tot += ((filterLo[j] * srcLo[k*srcwidth+arg] + filterHi[j] * srcHi[k*srcwidth+arg]));
|
||||
}
|
||||
dst[k*dstwidth+i] = tot;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
#ifdef __SSE2__
|
||||
template<typename T> SSEFUNCTION void wavelet_level<T>::SynthesisFilterSubsampVertical (T * RESTRICT srcLo, T * RESTRICT srcHi, T * RESTRICT dst, float (* RESTRICT filterLo)[4], float (* RESTRICT filterHi)[4], const int taps, const int offset, const int width, const int srcheight, const int dstheight)
|
||||
{
|
||||
|
||||
template<typename T> void wavelet_level<T>::SynthesisFilterSubsampVertical (T * RESTRICT srcLo, T * RESTRICT srcHi, T * RESTRICT dst, float * RESTRICT filterLo, float * RESTRICT filterHi, int taps, int offset, int pitch, int dstlen) {
|
||||
|
||||
/* Basic convolution code
|
||||
* Applies an FIR filter 'filter' with filter length 'taps',
|
||||
* aligning the 'offset' element of the filter with
|
||||
* the input pixel, and skipping 'skip' pixels between taps
|
||||
* Output is subsampled by two
|
||||
*/
|
||||
/* Basic convolution code
|
||||
* Applies an FIR filter 'filter' with filter length 'taps',
|
||||
* aligning the 'offset' element of the filter with
|
||||
* the input pixel, and skipping 'skip' pixels between taps
|
||||
* Output is subsampled by two
|
||||
*/
|
||||
|
||||
// calculate coefficients
|
||||
int srclen = (dstlen==m_w ? m_w2 : m_h2);//length of row/col in src (coarser level)
|
||||
int shift=skip*(taps-offset-1);//align filter with data
|
||||
|
||||
for(size_t i = m_pad; i < (dstlen+m_pad); i++) {
|
||||
__m128 fourv = _mm_set1_ps(4.f);
|
||||
for(size_t i = 0; i < dstheight; i++) {
|
||||
int i_src = (i+shift)/2;
|
||||
int begin = (i+shift)%2;
|
||||
//TODO: this is correct only if skip=1; otherwise, want to work with cosets of length 'skip'
|
||||
if (LIKELY(i>skip*taps && i<(srclen-skip*taps))) {//bulk
|
||||
for (int k=0; k<pitch; k++) {
|
||||
float tot = 0.f;
|
||||
if (LIKELY(i>skip*taps && i<(dstheight-skip*taps))) {//bulk
|
||||
int k;
|
||||
for (k=0; k<width-3; k+=4) {
|
||||
__m128 totv = _mm_setzero_ps();
|
||||
for (int j=begin, l=0; j<taps; j+=2, l+=skip) {
|
||||
tot += ((filterLo[j] * srcLo[(i_src-l)*pitch+k] + filterHi[j] * srcHi[(i_src-l)*pitch+k]));
|
||||
totv += ((LVF(filterLo[j][0]) * LVFU(srcLo[(i_src-l)*width+k]) + LVF(filterHi[j][0]) * LVFU(srcHi[(i_src-l)*width+k])));
|
||||
}
|
||||
dst[pitch*(i-m_pad)+k] = 4.f * tot;
|
||||
_mm_storeu_ps(&dst[width*i+k], fourv * totv);
|
||||
}
|
||||
} else {//boundary
|
||||
for (int k=0; k<pitch; k++) {
|
||||
for (; k<width; k++) {
|
||||
float tot = 0.f;
|
||||
for (int j=begin, l=0; j<taps; j+=2, l+=skip) {
|
||||
int arg = max(0,min((i_src-l),srclen-1))*pitch+k;//clamped BC's
|
||||
tot += ((filterLo[j] * srcLo[arg] + filterHi[j] * srcHi[arg]));
|
||||
tot += ((filterLo[j][0] * srcLo[(i_src-l)*width+k] + filterHi[j][0] * srcHi[(i_src-l)*width+k]));
|
||||
}
|
||||
dst[width*i+k] = 4.f * tot;
|
||||
}
|
||||
} else {//boundary
|
||||
int k;
|
||||
for (k=0; k<width-3; k+=4) {
|
||||
__m128 totv = _mm_setzero_ps();
|
||||
for (int j=begin, l=0; j<taps; j+=2, l+=skip) {
|
||||
int arg = max(0,min((i_src-l),srcheight-1))*width+k;//clamped BC's
|
||||
totv += ((LVF(filterLo[j][0]) * LVFU(srcLo[arg]) + LVF(filterHi[j][0]) * LVFU(srcHi[arg])));
|
||||
}
|
||||
dst[pitch*(i-m_pad)+k] = 4.f * tot;
|
||||
_mm_storeu_ps(&dst[width*i+k], fourv * totv);
|
||||
}
|
||||
for (; k<width; k++) {
|
||||
float tot = 0.f;
|
||||
for (int j=begin, l=0; j<taps; j+=2, l+=skip) {
|
||||
int arg = max(0,min((i_src-l),srcheight-1))*width+k;//clamped BC's
|
||||
tot += ((filterLo[j][0] * srcLo[arg] + filterHi[j][0] * srcHi[arg]));
|
||||
}
|
||||
dst[width*i+k] = 4.f * tot;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
#else
|
||||
template<typename T> void wavelet_level<T>::SynthesisFilterSubsampVertical (T * RESTRICT srcLo, T * RESTRICT srcHi, T * RESTRICT dst, float * RESTRICT filterLo, float * RESTRICT filterHi, const int taps, const int offset, const int width, const int srcheight, const int dstheight)
|
||||
{
|
||||
|
||||
template<typename T> template<typename E> void wavelet_level<T>::decompose_level(E *src, E *dst, float *filterV, float *filterH, int taps, int offset) {
|
||||
/* Basic convolution code
|
||||
* Applies an FIR filter 'filter' with filter length 'taps',
|
||||
* aligning the 'offset' element of the filter with
|
||||
* the input pixel, and skipping 'skip' pixels between taps
|
||||
* Output is subsampled by two
|
||||
*/
|
||||
|
||||
// calculate coefficients
|
||||
int shift=skip*(taps-offset-1);//align filter with data
|
||||
|
||||
for(size_t i = 0; i < dstheight; i++) {
|
||||
int i_src = (i+shift)/2;
|
||||
int begin = (i+shift)%2;
|
||||
//TODO: this is correct only if skip=1; otherwise, want to work with cosets of length 'skip'
|
||||
if (LIKELY(i>skip*taps && i<(dstheight-skip*taps))) {//bulk
|
||||
for (int k=0; k<width; k++) {
|
||||
float tot = 0.f;
|
||||
for (int j=begin, l=0; j<taps; j+=2, l+=skip) {
|
||||
tot += ((filterLo[j] * srcLo[(i_src-l)*width+k] + filterHi[j] * srcHi[(i_src-l)*width+k]));
|
||||
}
|
||||
dst[width*i+k] = 4.f * tot;
|
||||
}
|
||||
} else {//boundary
|
||||
for (int k=0; k<width; k++) {
|
||||
float tot = 0.f;
|
||||
for (int j=begin, l=0; j<taps; j+=2, l+=skip) {
|
||||
int arg = max(0,min((i_src-l),srcheight-1))*width+k;//clamped BC's
|
||||
tot += ((filterLo[j] * srcLo[arg] + filterHi[j] * srcHi[arg]));
|
||||
}
|
||||
dst[width*i+k] = 4.f * tot;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
#endif
|
||||
|
||||
#ifdef __SSE2__
|
||||
template<typename T> template<typename E> SSEFUNCTION void wavelet_level<T>::decompose_level(E *src, E *dst, float *filterV, float *filterH, int taps, int offset) {
|
||||
|
||||
T tmpLo[m_w] ALIGNED64;
|
||||
T tmpHi[m_w] ALIGNED64;
|
||||
/* filter along rows and columns */
|
||||
if(subsamp_out) {
|
||||
float filterVarray[2*taps][4] ALIGNED64;
|
||||
for(int i=0;i<2*taps;i++) {
|
||||
for(int j=0;j<4;j++) {
|
||||
filterVarray[i][j] = filterV[i];
|
||||
}
|
||||
}
|
||||
for(int row=0;row<m_h;row+=2) {
|
||||
AnalysisFilterSubsampVertical (src, tmpLo, tmpHi, filterV, filterV+taps, taps, offset, m_w/*output_pitch*/, m_h/*srclen*/, row);
|
||||
AnalysisFilterSubsampHorizontal (tmpLo, dst, wavcoeffs[1], filterH, filterH+taps, taps, offset, m_h2/*output_pitch*/, m_w/*srclen*/, m_w2, row/2);
|
||||
AnalysisFilterSubsampHorizontal (tmpHi, wavcoeffs[2], wavcoeffs[3], filterH, filterH+taps, taps, offset, m_h2/*output_pitch*/, m_w/*srclen*/, m_w2, row/2);
|
||||
AnalysisFilterSubsampVertical (src, tmpLo, tmpHi, filterVarray, filterVarray+taps, taps, offset, m_w, m_h, row);
|
||||
AnalysisFilterSubsampHorizontal (tmpLo, dst, wavcoeffs[1], filterH, filterH+taps, taps, offset, m_w, m_w2, row/2);
|
||||
AnalysisFilterSubsampHorizontal (tmpHi, wavcoeffs[2], wavcoeffs[3], filterH, filterH+taps, taps, offset, m_w, m_w2, row/2);
|
||||
}
|
||||
} else {
|
||||
for(int row=0;row<m_h;row++) {
|
||||
@@ -375,21 +519,64 @@ namespace rtengine {
|
||||
}
|
||||
}
|
||||
}
|
||||
#else
|
||||
template<typename T> template<typename E> void wavelet_level<T>::decompose_level(E *src, E *dst, float *filterV, float *filterH, int taps, int offset) {
|
||||
|
||||
T tmpLo[m_w] ALIGNED64;
|
||||
T tmpHi[m_w] ALIGNED64;
|
||||
/* filter along rows and columns */
|
||||
if(subsamp_out) {
|
||||
for(int row=0;row<m_h;row+=2) {
|
||||
AnalysisFilterSubsampVertical (src, tmpLo, tmpHi, filterV, filterV+taps, taps, offset, m_w, m_h, row);
|
||||
AnalysisFilterSubsampHorizontal (tmpLo, dst, wavcoeffs[1], filterH, filterH+taps, taps, offset, m_w, m_w2, row/2);
|
||||
AnalysisFilterSubsampHorizontal (tmpHi, wavcoeffs[2], wavcoeffs[3], filterH, filterH+taps, taps, offset, m_w, m_w2, row/2);
|
||||
}
|
||||
} else {
|
||||
for(int row=0;row<m_h;row++) {
|
||||
AnalysisFilterHaarVertical (src, tmpLo, tmpHi, m_w, m_h, row);
|
||||
AnalysisFilterHaarHorizontal (tmpLo, dst, wavcoeffs[1], m_w, row);
|
||||
AnalysisFilterHaarHorizontal (tmpHi, wavcoeffs[2], wavcoeffs[3], m_w, row);
|
||||
}
|
||||
}
|
||||
}
|
||||
#endif
|
||||
|
||||
#ifdef __SSE2__
|
||||
|
||||
template<typename T> template<typename E> SSEFUNCTION void wavelet_level<T>::reconstruct_level(E* tmpLo, E* tmpHi, E * src, E *dst, float *filterV, float *filterH, int taps, int offset) {
|
||||
|
||||
/* filter along rows and columns */
|
||||
if (subsamp_out) {
|
||||
float filterVarray[2*taps][4] ALIGNED64;
|
||||
for(int i=0;i<2*taps;i++) {
|
||||
for(int j=0;j<4;j++) {
|
||||
filterVarray[i][j] = filterV[i];
|
||||
}
|
||||
}
|
||||
SynthesisFilterSubsampHorizontal (src, wavcoeffs[1], tmpLo, filterH, filterH+taps, taps, offset, m_w2, m_w, m_h2);
|
||||
SynthesisFilterSubsampHorizontal (wavcoeffs[2], wavcoeffs[3], tmpHi, filterH, filterH+taps, taps, offset, m_w2, m_w, m_h2);
|
||||
SynthesisFilterSubsampVertical (tmpLo, tmpHi, dst, filterVarray, filterVarray+taps, taps, offset, m_w, m_h2, m_h);
|
||||
} else {
|
||||
SynthesisFilterHaarHorizontal (src, wavcoeffs[1], tmpLo, m_w, m_h2);
|
||||
SynthesisFilterHaarHorizontal (wavcoeffs[2], wavcoeffs[3], tmpHi, m_w, m_h2);
|
||||
SynthesisFilterHaarVertical (tmpLo, tmpHi, dst, m_w, m_h);
|
||||
}
|
||||
}
|
||||
#else
|
||||
template<typename T> template<typename E> void wavelet_level<T>::reconstruct_level(E* tmpLo, E* tmpHi, E * src, E *dst, float *filterV, float *filterH, int taps, int offset) {
|
||||
|
||||
/* filter along rows and columns */
|
||||
if (subsamp_out) {
|
||||
SynthesisFilterSubsampHorizontal (src, wavcoeffs[1], tmpLo, filterH, filterH+taps, taps, offset, m_w/*dstlen*/);
|
||||
SynthesisFilterSubsampHorizontal (wavcoeffs[2], wavcoeffs[3], tmpHi, filterH, filterH+taps, taps, offset, m_w/*dstlen*/);
|
||||
SynthesisFilterSubsampVertical (tmpLo, tmpHi, dst, filterV, filterV+taps, taps, offset, m_w/*pitch*/, m_h/*dstlen*/);
|
||||
SynthesisFilterSubsampHorizontal (src, wavcoeffs[1], tmpLo, filterH, filterH+taps, taps, offset, m_w2, m_w, m_h2);
|
||||
SynthesisFilterSubsampHorizontal (wavcoeffs[2], wavcoeffs[3], tmpHi, filterH, filterH+taps, taps, offset, m_w2, m_w, m_h2);
|
||||
SynthesisFilterSubsampVertical (tmpLo, tmpHi, dst, filterV, filterV+taps, taps, offset, m_w, m_h2, m_h);
|
||||
} else {
|
||||
SynthesisFilterHaarHorizontal (src, wavcoeffs[1], tmpLo, m_w);
|
||||
SynthesisFilterHaarHorizontal (wavcoeffs[2], wavcoeffs[3], tmpHi, m_w);
|
||||
SynthesisFilterHaarHorizontal (src, wavcoeffs[1], tmpLo, m_w, m_h2);
|
||||
SynthesisFilterHaarHorizontal (wavcoeffs[2], wavcoeffs[3], tmpHi, m_w, m_h2);
|
||||
SynthesisFilterHaarVertical (tmpLo, tmpHi, dst, m_w, m_h);
|
||||
}
|
||||
}
|
||||
|
||||
#endif
|
||||
};
|
||||
|
||||
#endif
|
||||
|
||||
Reference in New Issue
Block a user