liz/rawTherapee
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

findHotDeadPixels: further speedup and cleanup, #5531

Browse Source
This commit is contained in:
Ingo Weyrich
2019-11-21 15:37:55 +01:00
parent 3eb1d241c9
commit 0b5b9b0f3e
1 changed files with 45 additions and 122 deletions
Download Patch File Download Diff File Expand all files Collapse all files

167
rtengine/badpixels.cc
View File

@@ -22,13 +22,39 @@
#include "pixelsmap.h"
#include "rawimage.h"
#include "rawimagesource.h"
#define BENCHMARK
//#define BENCHMARK
#include "StopWatch.h"
namespace
{
unsigned fc(const unsigned int cfa[2][2], int r, int c) {
return cfa[r & 1][c & 1];
}
inline void sum5x5(const array2D<float>& in, int col, float &sum) {
#ifdef __SSE2__
// sum up 5*4 = 20 values using SSE
// 10 fabs function calls and 10 float additions with SSE
const vfloat sumv = (vabsf(LVFU(in[0][col])) + vabsf(LVFU(in[1][col]))) +
(vabsf(LVFU(in[2][col])) + vabsf(LVFU(in[3][col]))) +
vabsf(LVFU(in[4][col]));
// horizontally add the values and add the result to hfnbrave
sum += vhadd(sumv);
// add remaining 5 values of last column
sum += (fabsf(in[0][col + 4]) + fabsf(in[1][col + 4])) +
(fabsf(in[2][col + 4]) + fabsf(in[3][col + 4])) +
fabsf(in[4][col + 4]);
#else
// 25 fabs function calls and 25 float additions without SSE
for (int nn = col; nn < col + 5; ++nn) {
sum += (fabsf(in[0][nn]) + fabsf(in[1][nn])) +
(fabsf(in[2][nn]) + fabsf(in[3][nn])) +
fabsf(in[4][nn]);
}
#endif
}
}
namespace rtengine
@@ -446,8 +472,8 @@ int RawImageSource::interpolateBadPixelsXtrans(const PixelsMap &bitmapBads)
/* Search for hot or dead pixels in the image and update the map
* For each pixel compare its value to the average of similar color surrounding
* (Taken from Emil Martinec idea)
* (Optimized by Ingo Weyrich 2013, 2015 and 2019)
*/
* (Optimized by Ingo Weyrich 2013, 2015, and 2019)
*/
int RawImageSource::findHotDeadPixels(PixelsMap &bpMap, const float thresh, const bool findHotPixels, const bool findDeadPixels) const
{
BENCHFUN
@@ -460,11 +486,7 @@ int RawImageSource::findHotDeadPixels(PixelsMap &bpMap, const float thresh, cons
#pragma omp parallel reduction(+:counter)
#endif
{
array2D<float> cfablur(W, 5);
// zero left and right border
for (int i = 0; i < 5; ++i) {
cfablur[i][0] = cfablur[i][1] = cfablur[i][W - 2] = cfablur[i][W - 1];
}
array2D<float> cfablur(W, 5, ARRAY2D_CLEAR_DATA);
int firstRow = -1;
int lastRow = -1;
@@ -473,28 +495,13 @@ int RawImageSource::findHotDeadPixels(PixelsMap &bpMap, const float thresh, cons
#pragma omp for schedule(static) nowait
#endif
for (int i = 2; i < H - 2; i++) {
for (int i = 2; i < H - 2; ++i) {
if (firstRow == -1) {
firstRow = i;
if (firstRow == 2) {
for (int i = 0; i < 2; ++i) {
for (int j = 0; j < W; ++j) {
cfablur[i][j] = 0.f;
}
}
} else {
if (firstRow > 2) {
for (int row = firstRow - 2; row < firstRow; ++row) {
int destRow = row % 5;
int j = 2;
#ifdef __SSE2__
for (; j < W - 5; j += 4) {
const vfloat tempv = median(LVFU(rawData[row - 2][j - 2]), LVFU(rawData[row - 2][j]), LVFU(rawData[row - 2][j + 2]),
LVFU(rawData[row][j - 2]), LVFU(rawData[row][j]), LVFU(rawData[row][j + 2]),
LVFU(rawData[row + 2][j - 2]), LVFU(rawData[row + 2][j]), LVFU(rawData[row + 2][j + 2]));
STVFU(cfablur[destRow][j], LVFU(rawData[row][j]) - tempv);
}
#endif
for (; j < W - 2; j++) {
const int destRow = row % 5;
for (int j = 2; j < W - 2; ++j) {
const float temp = median(rawData[row - 2][j - 2], rawData[row - 2][j], rawData[row - 2][j + 2],
rawData[row][j - 2], rawData[row][j], rawData[row][j + 2],
rawData[row + 2][j - 2], rawData[row + 2][j], rawData[row + 2][j + 2]);
@@ -505,16 +512,7 @@ int RawImageSource::findHotDeadPixels(PixelsMap &bpMap, const float thresh, cons
}
lastRow = i;
const int destRow = i % 5;
int j = 2;
#ifdef __SSE2__
for (; j < W - 5; j += 4) {
const vfloat tempv = median(LVFU(rawData[i - 2][j - 2]), LVFU(rawData[i - 2][j]), LVFU(rawData[i - 2][j + 2]),
LVFU(rawData[i][j - 2]), LVFU(rawData[i][j]), LVFU(rawData[i][j + 2]),
LVFU(rawData[i + 2][j - 2]), LVFU(rawData[i + 2][j]), LVFU(rawData[i + 2][j + 2]));
STVFU(cfablur[destRow][j], LVFU(rawData[i][j]) - tempv);
}
#endif
for (; j < W - 2; j++) {
for (int j = 2; j < W - 2; ++j) {
const float temp = median(rawData[i - 2][j - 2], rawData[i - 2][j], rawData[i - 2][j + 2],
rawData[i][j - 2], rawData[i][j], rawData[i][j + 2],
rawData[i + 2][j - 2], rawData[i + 2][j], rawData[i + 2][j + 2]);
@@ -523,60 +521,26 @@ int RawImageSource::findHotDeadPixels(PixelsMap &bpMap, const float thresh, cons
if (i - 1 > firstRow) {
const int rr = i - 2;
const int rrm2 = (rr - 2) % 5;
const int rrm1 = (rr - 1) % 5;
const int rr0 = rr % 5;
const int rrp1 = (rr + 1) % 5;
const int rrp2 = (rr + 2) % 5;
for (int cc = 2; cc < W - 2; ++cc) {
//evaluate pixel for heat/death
float pixdev = cfablur[rr0][cc];
if (pixdev == 0.f) {
if (!findDeadPixels && pixdev <= 0.f) {
continue;
}
if ((!findDeadPixels) && pixdev < 0) {
continue;
}
if ((!findHotPixels) && pixdev > 0) {
if (!findHotPixels && pixdev >= 0.f) {
continue;
}
pixdev = fabsf(pixdev);
float hfnbrave = -pixdev;
#ifdef __SSE2__
// sum up 5*4 = 20 values using SSE
// 10 fabs function calls and 10 float additions with SSE
vfloat sum1 = vabsf(LVFU(cfablur[rrm2][cc - 2])) + vabsf(LVFU(cfablur[rrm1][cc - 2]));
vfloat sum2 = vabsf(LVFU(cfablur[rr0][cc - 2])) + vabsf(LVFU(cfablur[rrp1][cc - 2]));
sum1 += vabsf(LVFU(cfablur[rrp2][cc - 2]));
// horizontally add the values and add the result to hfnbrave
hfnbrave += vhadd(sum1 + sum2);
// add remaining 5 values of last column
hfnbrave += fabsf(cfablur[rrm2][cc + 2]);
hfnbrave += fabsf(cfablur[rrm1][cc + 2]);
hfnbrave += fabsf(cfablur[rr0][cc + 2]);
hfnbrave += fabsf(cfablur[rrp1][cc + 2]);
hfnbrave += fabsf(cfablur[rrp2][cc + 2]);
#else
// 25 fabs function calls and 25 float additions without SSE
for (int nn = cc - 2; nn <= cc + 2; ++nn) {
hfnbrave += fabsf(cfablur[rrm2][nn]);
hfnbrave += fabsf(cfablur[rrm1][nn]);
hfnbrave += fabsf(cfablur[rr0][nn]);
hfnbrave += fabsf(cfablur[rrp1][nn]);
hfnbrave += fabsf(cfablur[rrp2][nn]);
}
#endif
sum5x5(cfablur, cc - 2, hfnbrave);
if (pixdev > varthresh * hfnbrave) {
// mark the pixel as "bad"
bpMap.set(cc, rr);
counter++;
++counter;
}
} //end of pixel evaluation
}
@@ -592,16 +556,7 @@ int RawImageSource::findHotDeadPixels(PixelsMap &bpMap, const float thresh, cons
cfablur[destRow][j] = 0.f;
}
} else {
int j = 2;
#ifdef __SSE2__
for (; j < W - 5; j += 4) {
const vfloat tempv = median(LVFU(rawData[i - 2][j - 2]), LVFU(rawData[i - 2][j]), LVFU(rawData[i - 2][j + 2]),
LVFU(rawData[i][j - 2]), LVFU(rawData[i][j]), LVFU(rawData[i][j + 2]),
LVFU(rawData[i + 2][j - 2]), LVFU(rawData[i + 2][j]), LVFU(rawData[i + 2][j + 2]));
STVFU(cfablur[destRow][j], LVFU(rawData[i][j]) - tempv);
}
#endif
for (; j < W - 2; j++) {
for (int j = 2; j < W - 2; ++j) {
const float temp = median(rawData[i - 2][j - 2], rawData[i - 2][j], rawData[i - 2][j + 2],
rawData[i][j - 2], rawData[i][j], rawData[i][j + 2],
rawData[i + 2][j - 2], rawData[i + 2][j], rawData[i + 2][j + 2]);
@@ -609,64 +564,32 @@ int RawImageSource::findHotDeadPixels(PixelsMap &bpMap, const float thresh, cons
}
}
const int rrm2 = (rr - 2) % 5;
const int rrm1 = (rr - 1) % 5;
const int rr0 = rr % 5;
const int rrp1 = (rr + 1) % 5;
const int rrp2 = (rr + 2) % 5;
for (int cc = 2; cc < W - 2; ++cc) {
//evaluate pixel for heat/death
float pixdev = cfablur[rr0][cc];
if (pixdev == 0.f) {
if (!findDeadPixels && pixdev <= 0.f) {
continue;
}
if ((!findDeadPixels) && pixdev < 0) {
continue;
}
if ((!findHotPixels) && pixdev > 0) {
if (!findHotPixels && pixdev >= 0.f) {
continue;
}
pixdev = fabsf(pixdev);
float hfnbrave = -pixdev;
#ifdef __SSE2__
// sum up 5*4 = 20 values using SSE
// 10 fabs function calls and 10 float additions with SSE
vfloat sum1 = vabsf(LVFU(cfablur[rrm2][cc - 2])) + vabsf(LVFU(cfablur[rrm1][cc - 2]));
vfloat sum2 = vabsf(LVFU(cfablur[rr0][cc - 2])) + vabsf(LVFU(cfablur[rrp1][cc - 2]));
sum1 += vabsf(LVFU(cfablur[rrp2][cc - 2]));
// horizontally add the values and add the result to hfnbrave
hfnbrave += vhadd(sum1 + sum2);
// add remaining 5 values of last column
hfnbrave += fabsf(cfablur[rrm2][cc + 2]);
hfnbrave += fabsf(cfablur[rrm1][cc + 2]);
hfnbrave += fabsf(cfablur[rr0][cc + 2]);
hfnbrave += fabsf(cfablur[rrp1][cc + 2]);
hfnbrave += fabsf(cfablur[rrp2][cc + 2]);
#else
// 25 fabs function calls and 25 float additions without SSE
for (int nn = cc - 2; nn <= cc + 2; ++nn) {
hfnbrave += fabsf(cfablur[rrm2][nn]);
hfnbrave += fabsf(cfablur[rrm1][nn]);
hfnbrave += fabsf(cfablur[rr0][nn]);
hfnbrave += fabsf(cfablur[rrp1][nn]);
hfnbrave += fabsf(cfablur[rrp2][nn]);
}
#endif
sum5x5(cfablur, cc - 2, hfnbrave);
if (pixdev > varthresh * hfnbrave) {
// mark the pixel as "bad"
bpMap.set(cc, rr);
counter++;
++counter;
}
}//end of pixel evaluation
}
}
}//end of parallel processing
return counter;
}