From 2f2421442b5b8ef20290f214b0f8e9dd77076136 Mon Sep 17 00:00:00 2001
From: heckflosse <heckflosse67@gmx.de>
Date: Sun, 6 Mar 2016 01:14:35 +0100
Subject: [PATCH] Raw false colour suppression: additional 38% speedup

---
 rtengine/CMakeLists.txt     |   2 +-
 rtengine/improcfun.cc       |   3 -
 rtengine/impulse_denoise.h  |   1 +
 rtengine/median.h           | 232 +++++++++---------------------------
 rtengine/rawimagesource.cc  |  13 +-
 rtengine/rawimagesource_i.h |  10 +-
 6 files changed, 74 insertions(+), 187 deletions(-)

diff --git a/rtengine/CMakeLists.txt b/rtengine/CMakeLists.txt
index 597265bfe..1f4f3375f 100644
--- a/rtengine/CMakeLists.txt
+++ b/rtengine/CMakeLists.txt
@@ -11,7 +11,7 @@ set (RTENGINESOURCEFILES colortemp.cc curves.cc flatcurves.cc diagonalcurves.cc
     dfmanager.cc ffmanager.cc gauss.cc rawimage.cc image8.cc image16.cc imagefloat.cc imagedata.cc imageio.cc improcfun.cc init.cc dcrop.cc
     loadinitial.cc procparams.cc rawimagesource.cc demosaic_algos.cc shmap.cc simpleprocess.cc refreshmap.cc
     fast_demo.cc amaze_demosaic_RT.cc CA_correct_RT.cc cfa_linedn_RT.cc green_equil_RT.cc hilite_recon.cc expo_before_b.cc
-    stdimagesource.cc myfile.cc iccjpeg.cc hlmultipliers.cc improccoordinator.cc pipettebuffer.cc coord.cc
+    stdimagesource.cc myfile.cc iccjpeg.cc improccoordinator.cc pipettebuffer.cc coord.cc
     processingjob.cc rtthumbnail.cc utils.cc labimage.cc slicer.cc cieimage.cc
     iplab2rgb.cc ipsharpen.cc iptransform.cc ipresize.cc ipvibrance.cc
     imagedimensions.cc jpeg_memsrc.cc jdatasrc.cc iimage.cc
diff --git a/rtengine/improcfun.cc b/rtengine/improcfun.cc
index 339f5e07a..989d9c6a7 100644
--- a/rtengine/improcfun.cc
+++ b/rtengine/improcfun.cc
@@ -23,9 +23,6 @@
 #include "rtengine.h"
 #include "improcfun.h"
 #include "curves.h"
-#include "colorclip.h"
-#include "gauss.h"
-#include "bilateral2.h"
 #include "mytime.h"
 #include "iccstore.h"
 #include "impulse_denoise.h"
diff --git a/rtengine/impulse_denoise.h b/rtengine/impulse_denoise.h
index ba596559c..9d3b73b79 100644
--- a/rtengine/impulse_denoise.h
+++ b/rtengine/impulse_denoise.h
@@ -24,6 +24,7 @@
 #include "cieimage.h"
 #include "sleef.c"
 #include "opthelper.h"
+#include "gauss.h"
 
 using namespace std;
 
diff --git a/rtengine/median.h b/rtengine/median.h
index 4171bb299..c2c969492 100644
--- a/rtengine/median.h
+++ b/rtengine/median.h
@@ -27,181 +27,63 @@
   b3 = max(b3,a3);\
  }
 
-#define MERGESORT(a1,a2,a3,b1,b2,b3,c1,c2,c3,c4,c5,c6) \
-  {\
-  if (a1<b1) { \
-    c1 = a1; \
-    if (a2<b1) {  \
-      c2 = a2; \
-      if (a3<b1) { \
-        c3 = a3; c4 = b1; c5 = b2; c6 = b3;\
-      }\
-      else {\
-        c3 = b1;\
-        if (a3<b2) {\
-          c4 = a3; c5 = b2; c6 = b3;\
-        }\
-        else {\
-          c4 = b2;\
-          if (a3<b3) {\
-            c5 = a3; c6 = b3;\
-          }\
-          else {\
-            c5 = b3; c6 = a3;\
-          }\
-        }\
-      }\
-    }\
-    else {\
-      c2 = b1;\
-      if (a2<b2) {\
-        c3 = a2;\
-        if (a3<b2) {\
-          c4 = a3; c5 = b2; c6 = b3;\
-        }\
-        else {\
-          c4 = b2;\
-          if (a3<b3) {\
-            c5 = a3; c6 = b3;\
-          }\
-          else {\
-            c5 = b3; c6 = a3;\
-          }\
-        }\
-      }\
-      else {\
-        c3 = b2;\
-        if (a2<b3) {\
-          c4 = a2;\
-          if (a3<b3) {\
-            c5 = a3; c6 = b3;\
-          }\
-          else {\
-            c5 = b3; c6 = a3;\
-          }\
-        }\
-        else {\
-          c4 = b3; c5 = a2; c6 = a3;\
-        }\
-      }\
-    }\
-  }\
-  else {\
-    c1 = b1;\
-    if (a1<b2) {\
-      c2 = a1;\
-      if (a2<b2) {\
-        c3 = a2;\
-        if (a3<b2) {\
-          c4 = a3; c5 = b2; c6 = b3;\
-        }\
-        else {\
-          c4 = b2;\
-          if (a3<b3) {\
-            c5 = a3; c6 = b3;\
-          }\
-          else {\
-            c5 = b3; c6 = a3;\
-          }\
-        }\
-      }\
-      else {\
-        c3 = b2;\
-        if (a2<b3) {\
-          c4 = a2;\
-          if (a3<b3) {\
-            c5 = a3; c6 = b3;\
-          }\
-          else {\
-            c5 = b3; c6 = a3;\
-          }\
-        }\
-        else {\
-          c4 = b3; c5 = a2; c6 = a3;\
-        }\
-      }\
-    }\
-    else {\
-      c2 = b2;\
-      if (a1<b3) {\
-        c3 = a1;\
-        if (a2<b3) {\
-          c4 = a2;\
-          if (a3<b3) {\
-            c5 = a3; c6 = b3;\
-          }\
-          else {\
-            c5 = b3; c6 = a3;\
-          }\
-        }\
-        else {\
-          c4 = b3; c5 = a2; c6 = a3;\
-        }\
-      }\
-      else {\
-        c3 = b3; c4 = a1; c5 = a2; c6 = a3;\
-      }\
-    }\
-  }\
+
+#define NETWORKSORT4OF6(s0,s1,s2,s3,s4,s5,d0,d1,d2,d3,d4,d5,temp) \
+{\
+d1 = min(s1,s2);\
+d2 = max(s1,s2);\
+d0 = min(s0,d2);\
+d2 = max(s0,d2);\
+temp = min(d0,d1);\
+d1 = max(d0,d1);\
+d0 = temp;\
+d4 = min(s4,s5);\
+d5 = max(s4,s5);\
+temp = min(s3,d5);\
+d5 = max(s3,d5);\
+d3 = temp;\
+temp = min(d3,d4);\
+d4 = max(d3,d4);\
+d3 = temp;\
+d3 = max(d0,d3);\
+temp = min(d1,d4);\
+d4 = max(d1,d4);\
+d1 = temp;\
+d2 = min(d2,d5);\
+temp = min(d2,d4);\
+d4 = max(d2,d4);\
+d2 = temp;\
+temp = min(d1,d3);\
+d3 = max(d1,d3);\
+d1 = temp;\
+temp = min(d2,d3);\
+d3 = max(d2,d3);\
+d2 = temp;\
 }
 
-#define MEDIAN7(a1,a2,a3,b1,b2,b3,b4,median) \
-  { \
-  if (a1<b1) {\
-    if (a2<b1) {\
-      if (a3<b1) \
-        median = b1; \
-      else {\
-        if (a3<b2) \
-          median = a3;\
-        else \
-          median = b2;\
-      }\
-    }\
-    else {\
-      if (a2<b2) {\
-        if (a3<b2) \
-          median = a3;\
-        else \
-          median = b2;\
-      }\
-      else {\
-        if (a2<b3) \
-          median = a2;\
-        else \
-          median = b3;\
-      }\
-    }\
-  }\
-  else {\
-    if (a1<b2) {\
-      if (a2<b2) {\
-        if (a3<b2) \
-          median = a3;\
-        else\
-          median = b2;\
-      }\
-      else {\
-        if (a2<b3)\
-          median = a2;\
-        else\
-          median = b3;\
-      }\
-    }\
-    else {\
-      if (a1<b3) {\
-        if (a2<b3)\
-          median = a2;\
-        else \
-          median = b3; \
-      }\
-      else {\
-        if (a1<b4) \
-          median = a1;\
-        else\
-          median = b4;\
-      } \
-    }\
-  }\
+#define MEDIAN7(s0,s1,s2,s3,s4,s5,s6,t0,t1,t2,t3,t4,t5,t6,median) \
+{\
+t0 = min(s0,s5);\
+t5 = max(s0,s5);\
+t3 = max(t0,s3);\
+t0 = min(t0,s3);\
+t1 = min(s1,s6);\
+t6 = max(s1,s6);\
+t2 = min(s2,s4);\
+t4 = max(s2,s4);\
+t1 = max(t0,t1);\
+median = min(t3,t5);\
+t5 = max(t3,t5);\
+t3 = median;\
+median = min(t2,t6);\
+t6 = max(t2,t6);\
+t2 = median;\
+t3 = max(t2,t3);\
+t3 = min(t3,t6);\
+t4 = min(t4,t5);\
+median = min(t1,t4);\
+t4 = max(t1,t4);\
+t1 = median;\
+t3 = max(t1,t3);\
+median = min(t3,t4);\
 }
-
diff --git a/rtengine/rawimagesource.cc b/rtengine/rawimagesource.cc
index 75aab5951..dc7070c64 100644
--- a/rtengine/rawimagesource.cc
+++ b/rtengine/rawimagesource.cc
@@ -3437,12 +3437,13 @@ void RawImageSource::processFalseColorCorrectionThread  (Imagefloat* im, int row
         SORT3(rbconv_Q[px][1], rbconv_Q[cx][1], rbconv_Q[nx][1], pre2_Q[0], pre2_Q[1], pre2_Q[2]);
 
         // median I channel
+        float temp[7];
         for (int j = 1; j < W - 2; j += 2) {
             SORT3(rbconv_I[px][j + 1], rbconv_I[cx][j + 1], rbconv_I[nx][j + 1], post1_I[0], post1_I[1], post1_I[2]);
             SORT3(rbconv_I[px][j + 2], rbconv_I[cx][j + 2], rbconv_I[nx][j + 2], post2_I[0], post2_I[1], post2_I[2]);
-            MERGESORT(pre2_I[0], pre2_I[1], pre2_I[2], post1_I[0], post1_I[1], post1_I[2], middle_I[0], middle_I[1], middle_I[2], middle_I[3], middle_I[4], middle_I[5]);
-            MEDIAN7(pre1_I[0], pre1_I[1], pre1_I[2], middle_I[1], middle_I[2], middle_I[3], middle_I[4], rbout_I[cx][j]);
-            MEDIAN7(post2_I[0], post2_I[1], post2_I[2], middle_I[1], middle_I[2], middle_I[3], middle_I[4], rbout_I[cx][j + 1]);
+            NETWORKSORT4OF6(pre2_I[0], pre2_I[1], pre2_I[2], post1_I[0], post1_I[1], post1_I[2], middle_I[0], middle_I[1], middle_I[2], middle_I[3], middle_I[4], middle_I[5], temp[0]);
+            MEDIAN7(pre1_I[0], pre1_I[1], pre1_I[2], middle_I[1], middle_I[2], middle_I[3], middle_I[4], temp[0], temp[1], temp[2], temp[3], temp[4], temp[5], temp[6], rbout_I[cx][j]);
+            MEDIAN7(post2_I[0], post2_I[1], post2_I[2], middle_I[1], middle_I[2], middle_I[3], middle_I[4], temp[0], temp[1], temp[2], temp[3], temp[4], temp[5], temp[6], rbout_I[cx][j + 1]);
             tmp = pre1_I;
             pre1_I = post1_I;
             post1_I = tmp;
@@ -3456,9 +3457,9 @@ void RawImageSource::processFalseColorCorrectionThread  (Imagefloat* im, int row
         for (int j = 1; j < W - 2; j += 2) {
             SORT3(rbconv_Q[px][j + 1], rbconv_Q[cx][j + 1], rbconv_Q[nx][j + 1], post1_Q[0], post1_Q[1], post1_Q[2]);
             SORT3(rbconv_Q[px][j + 2], rbconv_Q[cx][j + 2], rbconv_Q[nx][j + 2], post2_Q[0], post2_Q[1], post2_Q[2]);
-            MERGESORT(pre2_Q[0], pre2_Q[1], pre2_Q[2], post1_Q[0], post1_Q[1], post1_Q[2], middle_Q[0], middle_Q[1], middle_Q[2], middle_Q[3], middle_Q[4], middle_Q[5]);
-            MEDIAN7(pre1_Q[0], pre1_Q[1], pre1_Q[2], middle_Q[1], middle_Q[2], middle_Q[3], middle_Q[4], rbout_Q[cx][j]);
-            MEDIAN7(post2_Q[0], post2_Q[1], post2_Q[2], middle_Q[1], middle_Q[2], middle_Q[3], middle_Q[4], rbout_Q[cx][j + 1]);
+            NETWORKSORT4OF6(pre2_Q[0], pre2_Q[1], pre2_Q[2], post1_Q[0], post1_Q[1], post1_Q[2], middle_Q[0], middle_Q[1], middle_Q[2], middle_Q[3], middle_Q[4], middle_Q[5], temp[0]);
+            MEDIAN7(pre1_Q[0], pre1_Q[1], pre1_Q[2], middle_Q[1], middle_Q[2], middle_Q[3], middle_Q[4], temp[0], temp[1], temp[2], temp[3], temp[4], temp[5], temp[6], rbout_Q[cx][j]);
+            MEDIAN7(post2_Q[0], post2_Q[1], post2_Q[2], middle_Q[1], middle_Q[2], middle_Q[3], middle_Q[4], temp[0], temp[1], temp[2], temp[3], temp[4], temp[5], temp[6], rbout_Q[cx][j + 1]);
             tmp = pre1_Q;
             pre1_Q = post1_Q;
             post1_Q = tmp;
diff --git a/rtengine/rawimagesource_i.h b/rtengine/rawimagesource_i.h
index 8ef2e435f..83e834924 100644
--- a/rtengine/rawimagesource_i.h
+++ b/rtengine/rawimagesource_i.h
@@ -22,13 +22,15 @@
 
 #include "rawimagesource.h"
 
-#include "curves.h"
-
 namespace rtengine
 {
 
 inline void RawImageSource::convert_row_to_YIQ (const float* const r, const float* const g, const float* const b, float* Y, float* I, float* Q, const int W)
 {
+#ifdef _OPENMP
+    #pragma omp simd
+#endif
+
     for (int j = 0; j < W; j++) {
         Y[j] = .299f * r[j] + .587f * g[j] + .114f * b[j];
         I[j] = .596f * r[j] - .275f * g[j] - .321f * b[j];
@@ -38,6 +40,10 @@ inline void RawImageSource::convert_row_to_YIQ (const float* const r, const floa
 
 inline void RawImageSource::convert_row_to_RGB (float* r, float* g, float* b, const float* const Y, const float* const I, const float* const Q, const int W)
 {
+#ifdef _OPENMP
+    #pragma omp simd
+#endif
+
     for (int j = 1; j < W - 1; j++) {
         r[j] = Y[j] + 0.956f * I[j] + 0.621f * Q[j];
         g[j] = Y[j] - 0.272f * I[j] - 0.647f * Q[j];