From 906cd4ba3af3de9677e13688d1eb27f713a1a2b7 Mon Sep 17 00:00:00 2001
From: Ingo <heckflosse@i-weyrich.de>
Date: Wed, 5 Feb 2014 23:09:55 +0100
Subject: [PATCH] Optimization of RawImage::get_colorsCoeff, Issue 2227

---
 rtengine/rawimage.cc | 109 +++++++++++++++++++++++++++++++++----------
 1 file changed, 84 insertions(+), 25 deletions(-)

diff --git a/rtengine/rawimage.cc b/rtengine/rawimage.cc
index 0a8d17d0e..acc7960d6 100755
--- a/rtengine/rawimage.cc
+++ b/rtengine/rawimage.cc
@@ -47,9 +47,8 @@ RawImage::~RawImage()
 /* Similar to dcraw scale_colors for coeff. calculation, but without actual pixels scaling.
  * need pixels in data[][] available
  */
-void RawImage::get_colorsCoeff( float *pre_mul_, float *scale_mul_, float *cblack_, bool forceAutoWB)
+void RawImage::get_colorsCoeff( float *pre_mul_, float *scale_mul_, float *cblack_, bool forceAutoWB) {
 
-{
 	unsigned  row, col, x, y, c, sum[8];
 	unsigned  W = this->get_width();
 	unsigned  H = this->get_height();
@@ -71,30 +70,90 @@ void RawImage::get_colorsCoeff( float *pre_mul_, float *scale_mul_, float *cblac
 	}
 	if ( this->get_cam_mul(0) == -1 || forceAutoWB) {
 		memset(dsum, 0, sizeof dsum);
-		for (row = 0; row < H; row += 8)
-			for (col = 0; col < W ; col += 8) {
-				memset(sum, 0, sizeof sum);
-				for (y = row; y < row + 8 && y < H; y++)
-					for (x = col; x < col + 8 && x < W; x++)
-						for (int c = 0; c < 3; c++) {
-							if (this->isBayer()) {
-								c = FC(y, x);
-								val = data[y][x];
-							} else
-								val = data[y][3*x+c];
-							if (val > this->get_white(c) - 25)
-								goto skip_block;
-							if ((val -= cblack_[c]) < 0)
-								val = 0;
-							sum[c] += val;
-							sum[c + 4]++;
-							if ( this->isBayer())
-								break;
-						}
-				for (c = 0; c < 8; c++)
-					dsum[c] += sum[c];
-skip_block: ;
+		if (this->isBayer()) {
+			// calculate number of pixels per color
+			dsum[FC(0,0)+4] += (int)(((W+1)/2) * ((H+1)/2));
+			dsum[FC(0,1)+4] += (int)(((W/2) * ((H+1)/2)));
+			dsum[FC(1,0)+4] += (int)(((W+1)/2) * (H/2));
+			dsum[FC(1,1)+4] += (int)((W/2) * (H/2));
+
+#pragma omp parallel private(val,row,col,x,y)
+{
+			double dsumthr[8];
+			memset(dsumthr, 0, sizeof dsumthr);
+			int sum[4];
+			// make local copies of the black and white values to avoid calculations and conversions
+			int cblackint[4];
+			int whiteint[4];
+			for (int c = 0; c < 4; c++) {
+				cblackint[c] = cblack_[c];
+				whiteint[c] = this->get_white(c) - 25;
 			}
+			unsigned short *tempdata = data[0];
+#pragma omp for nowait
+			for (row = 0; row < H; row += 8) {
+				int ymax = row + 8 < H ? row + 8 : H;
+				for (col = 0; col < W ; col += 8) {
+					int xmax = col + 8 < W ? col + 8 : W;
+					memset(sum, 0, sizeof sum);
+					for (y = row; y < ymax; y++)
+						for (x = col; x < xmax; x++) {
+							int c = FC(y, x);
+							val = tempdata[y*W+x];
+							if (val > whiteint[c]) { // calculate number of pixels to be substracted from sum and skip the block
+								dsumthr[FC(row,col)+4]		+= (int)(((xmax - col + 1)/2) * ((ymax - row + 1)/2));
+								dsumthr[FC(row,col+1)+4]	+= (int)(((xmax - col)/2) * ((ymax - row + 1)/2));
+								dsumthr[FC(row+1,col)+4]	+= (int)(((xmax - col + 1)/2) * ((ymax - row)/2));
+								dsumthr[FC(row+1,col+1)+4]	+= (int)(((xmax - col)/2) * ((ymax - row)/2));
+								goto skip_block2;
+							}
+							if (val < cblackint[c])
+								val = cblackint[c];
+							sum[c] += val;
+						}
+					for (int c = 0; c < 4; c++)
+						dsumthr[c] += sum[c];
+skip_block2: ;
+				}
+			}
+#pragma omp critical
+{
+			for (int c = 0; c < 4; c++)
+				dsum[c] += dsumthr[c];
+			for (int c = 4; c < 8; c++)
+				dsum[c] -= dsumthr[c];
+	
+}
+}
+			for(int c=0;c<4;c++)
+				dsum[c] -= cblack_[c] * dsum[c+4];
+
+		} else {
+			for (row = 0; row < H; row += 8)
+				for (col = 0; col < W ; col += 8) {
+					memset(sum, 0, sizeof sum);
+					for (y = row; y < row + 8 && y < H; y++)
+						for (x = col; x < col + 8 && x < W; x++)
+							for (int c = 0; c < 3; c++) {
+								if (this->isBayer()) {
+									c = FC(y, x);
+									val = data[y][x];
+								} else
+									val = data[y][3*x+c];
+								if (val > this->get_white(c) - 25)
+									goto skip_block;
+								if ((val -= cblack_[c]) < 0)
+									val = 0;
+								sum[c] += val;
+								sum[c + 4]++;
+								if ( this->isBayer())
+									break;
+							}
+					for (c = 0; c < 8; c++)
+						dsum[c] += sum[c];
+skip_block: ;
+				}
+		}
 		for (int c = 0; c < 4; c++)
 			if (dsum[c])
 				pre_mul_[c] = dsum[c + 4] / dsum[c];