Merge pull request #5236 from Beep6581/newlocallab-speedup

Speedups and cleanups for newlocallab

rtengine/improcfun.h

@@ -242,7 +242,7 @@ public:
 
     void luminanceCurve(LabImage* lold, LabImage* lnew, LUTf &curve);
 
-    void ciecamloc_02float(int sp, LabImage* lab, LabImage* dest);
+    void ciecamloc_02float(int sp, LabImage* lab);
 
     void ciecam_02float(CieImage* ncie, float adap, int pW, int pwb, LabImage* lab, const ProcParams* params,
                         const ColorAppearance & customColCurve1, const ColorAppearance & customColCurve, const ColorAppearance & customColCurve3,
@@ -250,7 +250,7 @@ public:
                         bool showSharpMask = false);
     void chromiLuminanceCurve(PipetteBuffer *pipetteBuffer, int pW, LabImage* lold, LabImage* lnew, LUTf &acurve, LUTf &bcurve, LUTf & satcurve, LUTf & satclcurve, LUTf &clcurve, LUTf &curve, bool utili, bool autili, bool butili, bool ccutili, bool cclutili, bool clcutili, LUTu &histCCurve, LUTu &histLurve);
     void vibrance(LabImage* lab);         //Jacques' vibrance
-    void softprocess(LabImage* bufcolorig, float ** buflight, /* float ** bufchro, float ** buf_a, float ** buf_b, */ float rad, int bfh, int bfw, int sk, bool multiThread);
+    void softprocess(const LabImage* bufcolorig, array2D<float> &buflight, /* float ** bufchro, float ** buf_a, float ** buf_b, */ float rad, int bfh, int bfw, int sk, bool multiThread);
 //    void colorCurve       (LabImage* lold, LabImage* lnew);
     void sharpening(LabImage* lab, const procparams::SharpeningParams &sharpenParam, bool showMask = false);
     void sharpeningcam(CieImage* ncie, float** buffer, bool showMask = false);
@@ -311,7 +311,7 @@ public:
     void vibrancelocal(int sp, int bfw, int bfh, LabImage* lab,  LabImage* dest, bool & localskutili, LUTf & sklocalcurve);
     void transit_shapedetect(int senstype, LabImage * bufexporig, LabImage * originalmask, float **buflight, float **bufchro, float **buf_a_cat, float ** buf_b_cat, float ** bufhh, bool HHutili, const float hueref, const float chromaref, const float lumaref, float sobelref, float meansobel, float ** blend2, const struct local_params & lp, LabImage * original, LabImage * transformed, int cx, int cy, int sk);
     void exlabLocal(const local_params& lp, int bfh, int bfw, LabImage* bufexporig, LabImage* lab,  LUTf & hltonecurve, LUTf & shtonecurve, LUTf & tonecurve);
-    void Exclude_Local(int sen, float **deltaso, const float hueref, const float chromaref, const float lumaref, float sobelref, float meansobel, const struct local_params & lp, LabImage * original, LabImage * transformed, LabImage * rsv, LabImage * reserv, int cx, int cy, int sk);
+    void Exclude_Local(float **deltaso, float hueref, float chromaref, float lumaref, float sobelref, float meansobel, const struct local_params & lp, const LabImage * original, LabImage * transformed, const LabImage * rsv, const LabImage * reserv, int cx, int cy, int sk);
 
     void DeNoise_Local(int call, const struct local_params& lp, int levred, float hueref, float lumaref, float chromaref, LabImage* original, LabImage* transformed, LabImage &tmp1, int cx, int cy, int sk);
 

rtengine/labimage.cc

@@ -25,9 +25,12 @@
 namespace rtengine
 {
 
-LabImage::LabImage (int w, int h) : W(w), H(h)
+LabImage::LabImage (int w, int h, bool initZero, bool multiThread) : W(w), H(h)
 {
     allocLab(w, h);
+    if (initZero) {
+        clear(multiThread);
+    }
 }
 
 LabImage::~LabImage ()
@@ -37,7 +40,19 @@ LabImage::~LabImage ()
 
 void LabImage::CopyFrom(LabImage *Img)
 {
+#ifdef _OPENMP
+    #pragma omp parallel sections
+    {
+        #pragma omp section
+        memcpy(L[0], Img->L[0], W * H * sizeof(float));
+        #pragma omp section
+        memcpy(a[0], Img->a[0], W * H * sizeof(float));
+        #pragma omp section
+        memcpy(b[0], Img->b[0], W * H * sizeof(float));
+    }
+#else
     memcpy(data, Img->data, W * H * 3 * sizeof(float));
+#endif
 }
 
 void LabImage::getPipetteData (float &v1, float &v2, float &v3, int posX, int posY, int squareSize)
@@ -107,12 +122,8 @@ void LabImage::clear(bool multiThread) {
 #ifdef _OPENMP
         #pragma omp parallel for if(multiThread)
 #endif
-        for(int i = 0; i < H; ++i) {
-            for(int j = 0; j < W; ++j) {
-                L[i][j] = a[i][j] = b[i][j] = 0.f;
-            }
+        for(size_t i = 0; i < static_cast<size_t>(H) * W * 3; ++i) {
+            data[i] = 0.f;
         }
     }
-
-
 }

rtengine/labimage.h

@@ -34,7 +34,7 @@ public:
     float** a;
     float** b;
 
-    LabImage (int w, int h);
+    LabImage (int w, int h, bool initZero = false, bool multiThread = true);
     ~LabImage ();
 
     //Copies image data in Img into this instance.