trying group phase correlation, preparing to convert to GPU

src/main/java/com/elphel/imagej/cameras/CLTParameters.java

@@ -32,7 +32,7 @@ public class CLTParameters {
 	public int        ishift_y =            0;  // debug feature - shift source image by this pixels down
 
 	private double    fat_zero =          0.05; // modify phase correlation to prevent division by very small numbers
-	private double    fat_zero_mono =     0.1;  // modify phase correlation to prevent division by very small numbers
+	private double    fat_zero_mono =     0.03;  // modify phase correlation to prevent division by very small numbers
 	private double    corr_sigma =        0.8;  // LPF correlation sigma
 	private double    corr_sigma_mono =   0.1;  // LPF correlation sigma for monochrome images
 	private double    scale_strength_main = 1.0; // leave as is
@@ -766,9 +766,10 @@ public class CLTParameters {
 	public boolean    taEnMismatch         = false;  // Enable cost of a measurement layer not having same layer in the same location or near
 
 // gpu processing parameters
+	public double     gpu_corr_scale =    0.75; // reduce GPU-generated correlation values
 	public int        gpu_corr_rad =        7;  // size of the correlation to save - initially only 15x15
-	public double     gpu_weight_r =      0.25;
-	public double     gpu_weight_b =      0.25; // weight g = 1.0 - gpu_weight_r - gpu_weight_b
+	public double     gpu_weight_r =      0.5; // 25;
+	public double     gpu_weight_b =      0.2; // 0.25; // weight g = 1.0 - gpu_weight_r - gpu_weight_b
 	public double     gpu_sigma_r =       0.9; // 1.1;
 	public double     gpu_sigma_b =       0.9; // 1.1;
 	public double     gpu_sigma_g =       0.6; // 0.7;
@@ -1567,6 +1568,7 @@ public class CLTParameters {
 		properties.setProperty(prefix+"taEnMismatch",               this.taEnMismatch +"");
 
 
+		properties.setProperty(prefix+"gpu_corr_scale",             this.gpu_corr_scale +"");
 		properties.setProperty(prefix+"gpu_corr_rad",               this.gpu_corr_rad +"");
 		properties.setProperty(prefix+"gpu_weight_r",               this.gpu_weight_r +"");
 		properties.setProperty(prefix+"gpu_weight_b",               this.gpu_weight_b +"");
@@ -2352,6 +2354,7 @@ public class CLTParameters {
 		if (properties.getProperty(prefix+"taEnFlaps")!=null)                   this.taEnFlaps=Boolean.parseBoolean(properties.getProperty(prefix+"taEnFlaps"));
 		if (properties.getProperty(prefix+"taEnMismatch")!=null)                this.taEnMismatch=Boolean.parseBoolean(properties.getProperty(prefix+"taEnMismatch"));
 
+		if (properties.getProperty(prefix+"gpu_corr_scale")!=null)              this.gpu_corr_scale=Double.parseDouble(properties.getProperty(prefix+"gpu_corr_scale"));
 		if (properties.getProperty(prefix+"gpu_corr_rad")!=null)                this.gpu_corr_rad=Integer.parseInt(properties.getProperty(prefix+"gpu_corr_rad"));
 		if (properties.getProperty(prefix+"gpu_weight_r")!=null)                this.gpu_weight_r=Double.parseDouble(properties.getProperty(prefix+"gpu_weight_r"));
 		if (properties.getProperty(prefix+"gpu_weight_b")!=null)                this.gpu_weight_b=Double.parseDouble(properties.getProperty(prefix+"gpu_weight_b"));
@@ -3279,6 +3282,8 @@ public class CLTParameters {
 
 		gd.addTab         ("GPU", "Parameters for GPU development");
 		gd.addMessage     ("--- GPU processing parameters ---");
+		gd.addNumericField("GPU 2D correlation scale",                                                                  this.gpu_corr_scale, 4, 6,"",
+				"Reduce GPU-generated correlation values to approximately match CPU-generated ones");
 		gd.addNumericField("Correlation radius",                                                                        this.gpu_corr_rad, 0, 6,"pix",
 				"Size of the 2D correlation - maximal radius = 7 corresponds to full 15x15 pixel tile");
 		gd.addNumericField("Correlation weight R",                                                                      this.gpu_weight_r, 4, 6,"",
@@ -4057,6 +4062,7 @@ public class CLTParameters {
 		this.taEnFlaps=             gd.getNextBoolean();
 		this.taEnMismatch=          gd.getNextBoolean();
 
+		this.gpu_corr_scale =       gd.getNextNumber();
 		this.gpu_corr_rad =   (int) gd.getNextNumber();
 		this.gpu_weight_r =         gd.getNextNumber();
 		this.gpu_weight_b =         gd.getNextNumber();

src/main/java/com/elphel/imagej/gpu/GPUTileProcessor.java

@@ -803,7 +803,7 @@ public class GPUTileProcessor {
 
 /**
  * Copy array of CPU-prepared tasks to the GPU memory
- * @param tile_tasks array of TpTask prepared by the CPU (before geometry correction is appllied)
+ * @param tile_tasks array of TpTask prepared by the CPU (before geometry correction is applied)
  * @param use_aux Use second (aux) camera
  */
         public void setTasks(TpTask [] tile_tasks, boolean use_aux) // while is it in class member? - just to be able to free
@@ -2078,6 +2078,16 @@ public class GPUTileProcessor {
     		if (debug) System.out.println("constantMemorySize: " + constantMemorySize);
             cuMemcpyHtoD(constantMemoryPointer, Pointer.to(lpf_flat), constantMemorySize);
             if (debug) System.out.println();
+            /*
+            if (debug) {
+        		for (int i = 0; i < lpf_flat.length; i++) {
+        			System.out.print(String.format("%8.5f", lpf_flat[i]));
+        			if (((i+1) % 16) == 0) {
+        				System.out.println();
+        			}
+        		}
+            }
+            */
     	}
 
     	public void setLpfCorr(
@@ -2094,6 +2104,16 @@ public class GPUTileProcessor {
     		if (debug) System.out.println("constantMemorySize: " + constantMemorySize);
             cuMemcpyHtoD(constantMemoryPointer, Pointer.to(lpf_flat), constantMemorySize);
             if (debug) System.out.println();
+            /*
+            if (debug) {
+        		for (int i = 0; i < lpf_flat.length; i++) {
+        			System.out.print(String.format("%8.5f", lpf_flat[i]));
+        			if (((i+1) % 16) == 0) {
+        				System.out.println();
+        			}
+        		}
+            }
+            */
     	}
 
     	public float [] floatSetCltLpfFd(

src/main/java/com/elphel/imagej/tileprocessor/ImageDtt.java

@@ -72,6 +72,7 @@ public class ImageDtt extends ImageDttCPU {
 			final float [][][]        iclt_fimg,       // will return quad images or null to skip, use quadCLT.linearStackToColor 
 ////		final int                 width,
 			// new parameters, will replace some other?
+			final double              gpu_corr_scale,  //  0.75; // reduce GPU-generated correlation values
 			final double              gpu_fat_zero,    // clt_parameters.getGpuFatZero(is_mono);absolute == 30.0
 			final double              gpu_sigma_r,     // 0.9, 1.1
 			final double              gpu_sigma_b,     // 0.9, 1.1
@@ -163,6 +164,7 @@ public class ImageDtt extends ImageDttCPU {
 		}
 */
 		// keep for now for mono, find out  what do they mean for macro mode
+		
 		if (macro_mode) { // all the same as they now mean different
 			//compensating Bayer correction
 			col_weights[0] = 0.25; //  1.0/3;
@@ -179,7 +181,7 @@ public class ImageDtt extends ImageDttCPU {
 				col_weights[1] = corr_blue * col_weights[2];
 			}
 		}
-
+/*
 		double [] scales = isMonochrome() ?
 				(new double [] {1.0}) :
 					(macro_mode?
@@ -189,13 +191,14 @@ public class ImageDtt extends ImageDttCPU {
 										corr_blue, // 0.25
 										1.0 - corr_red - corr_blue})); // 0.5
 		
-		
+*/		
 		final int corr_size = transform_size * 2 - 1;
-		final int [][] transpose_indices = new int [corr_size*(corr_size-1)/2][2];
+//		final int [][] transpose_indices = new int [corr_size*(corr_size-1)/2][2];
 		
 		if ((globalDebugLevel > -10) && (disparity_corr != 0.0)){
 			System.out.println(String.format("Using manual infinity disparity correction of %8.5f pixels",disparity_corr));
 		}
+/*		
 		{ int indx = 0;
 		for (int i =0; i < corr_size-1; i++){
 			for (int j = i+1; j < corr_size; j++){
@@ -204,7 +207,7 @@ public class ImageDtt extends ImageDttCPU {
 			}
 		}
 		}
-
+*/
 
 ////		final int first_color = isMonochrome()? MONO_CHN : 0; // color that is non-zero
 
@@ -316,8 +319,8 @@ public class ImageDtt extends ImageDttCPU {
 
 		DttRad2 dtt = new DttRad2(transform_size);
 		dtt.set_window(window_type);
-		final double [] lt_window = dtt.getWin2d();	// [256]
-		final double [] lt_window2 = new double [lt_window.length]; // squared
+		final double [] lt_window = dtt.getWin2d();	// [256] - never used
+		final double [] lt_window2 = new double [lt_window.length]; // squared - never used
 
 		for (int i = 0; i < lt_window.length; i++) lt_window2[i] = lt_window[i] * lt_window[i];
 
@@ -356,20 +359,20 @@ public class ImageDtt extends ImageDttCPU {
 		};
 		gpuQuad.setLpfRbg( // constants memory - same for all cameras
 				lpf_rgb,
-				false); // !batch_mode);
+				globalDebugLevel > -1);
 
 		final float [] lpf_flat = floatGetCltLpfFd(gpu_sigma_corr);
 
 		gpuQuad.setLpfCorr(// constants memory - same for all cameras
 				"lpf_corr", // String const_name, // "lpf_corr"
 				lpf_flat,
-				false); // !batch_mode);
+				globalDebugLevel > -1);
 
 		final float [] lpf_rb_flat = floatGetCltLpfFd(gpu_sigma_rb_corr);
 		gpuQuad.setLpfCorr(// constants memory - same for all cameras
 				"lpf_rb_corr", // String const_name, // "lpf_corr"
 				lpf_rb_flat,
-				false); // !batch_mode);
+				globalDebugLevel > -1);
 
 		gpuQuad.setTasks( // copy tp_tasks to the GPU memory
 				tp_tasks, // TpTask [] tile_tasks,
@@ -382,7 +385,7 @@ public class ImageDtt extends ImageDttCPU {
 			for (int ncam = 0; ncam < iclt_fimg.length; ncam++) {
 				iclt_fimg[ncam] = gpuQuad.getRBG(ncam); // retrieve data from GPU
 			}
-		}
+		} else {gpuQuad.execImcltRbgAll(isMonochrome());} // just for testing
 		// does it need texture tiles to be output?
 		if (texture_img != null) {
 			Rectangle woi = new Rectangle(); // will be filled out to match actual available image
@@ -443,7 +446,7 @@ public class ImageDtt extends ImageDttCPU {
 		if (fneed_corr) {
 			//Generate 2D phase correlations from the CLT representation
 			gpuQuad.execCorr2D(
-					scales,// double [] scales,
+					col_weights, // scales,// double [] scales,
 					gpu_fat_zero, // double fat_zero);
 					gpu_corr_rad); // int corr_radius
 			//Show 2D correlations
@@ -452,74 +455,43 @@ public class ImageDtt extends ImageDttCPU {
 			final float [][] fcorr2D = gpuQuad.getCorr2D(
 					gpu_corr_rad); //  int corr_rad);
 			if (corr_indices.length > 0) {
-				final int corr_length = fcorr2D[0].length;// all correlation tiles have the same size
-				// assuming that the correlation pairs sets are the same for each tile that has correlations
-				// find this number
-				int nt0 = (corr_indices[0] >> GPUTileProcessor.CORR_NTILE_SHIFT);
-				int nc0 = 1;
-				for (int i = 1; (i < corr_indices.length) && ((corr_indices[i] >> GPUTileProcessor.CORR_NTILE_SHIFT) == nt0) ; i++) {
-					nc0++;
-				}
-				final int num_tile_corr = nc0;
-				final int num_tiles = corr_indices.length / num_tile_corr; 
-				
-
-				/*
-
-			// convert to 6-layer image		 using tasks
+				if (true) {
+					int [] wh = new int[2];
 					double [][] dbg_corr = GPUTileProcessor.getCorr2DView(
 							tilesX,
 							tilesY,
 							corr_indices,
-					corr2D,
+							fcorr2D,
 							wh);
 					(new ShowDoubleFloatArrays()).showArrays(
 							dbg_corr,
 							wh[0],
 							wh[1],
 							true,
-						name+"-CORR2D-D"+clt_parameters.disparity,
+							"dbg-corr2D", // name+"-CORR2D-D"+clt_parameters.disparity,
 							GPUTileProcessor.getCorrTitles());
-				 */
+				}
+				
+				
+				final int corr_length = fcorr2D[0].length;// all correlation tiles have the same size
+				// assuming that the correlation pairs sets are the same for each tile that has correlations
+				// find this number
+				int nt0 = (corr_indices[0] >> GPUTileProcessor.CORR_NTILE_SHIFT);
+				int nc0 = 1;
+				for (int i = 1; (i < corr_indices.length) && ((corr_indices[i] >> GPUTileProcessor.CORR_NTILE_SHIFT) == nt0) ; i++) {
+					nc0++;
+				}
+				final int num_tile_corr = nc0;
+				final int num_tiles = corr_indices.length / num_tile_corr; 
+				
 
 				for (int ithread = 0; ithread < threads.length; ithread++) {
 					threads[ithread] = new Thread() {
 						@Override
 						public void run() {
 							//							int tileY,tileX,tIndex;
-							double [][]  corrs = new double [GPUTileProcessor.NUM_PAIRS][corr_length]; // 225-long (15x15)
-							
-/*
-					DttRad2 dtt = new DttRad2(transform_size);
-					dtt.set_window(window_type);
-					int tileY,tileX,tIndex; // , chn;
-					//						showDoubleFloatArrays sdfa_instance = new showDoubleFloatArrays(); // just for debugging?
-					double centerX; // center of aberration-corrected (common model) tile, X
-					double centerY; //
-					double [][] fract_shiftsXY = new double[quad][];
-					double [][]     tcorr_combo =    null; // [15*15] pixel space
-					double [][][]   tcorr_partial =  null; // [quad][numcol+1][15*15]
-					double [][][][] tcorr_tpartial = null; // [quad][numcol+1][4][8*8]
-					double [] ports_rgb = null;
-					double [][] rXY;
-
-					if (use_main) {
-						rXY = geometryCorrection.getRXY(true); // boolean use_rig_offsets,
-					} else  {
-						rXY = geometryCorrection.getRXY(false); // boolean use_rig_offsets,
-					}
-					Correlation2d corr2d = new Correlation2d(
-							imgdtt_params,              // ImageDttParameters  imgdtt_params,
-							transform_size,             // int transform_size,
-							2.0,                        //  double wndx_scale, // (wndy scale is always 1.0)
-							isMonochrome(), // boolean monochrome,
-							(globalDebugLevel > -1));   //   boolean debug)
-					corr2d.createOrtoNotch(
-							imgdtt_params.getEnhOrthoWidth(isAux()), // double getEnhOrthoWidth(isAux()),
-							imgdtt_params.getEnhOrthoScale(isAux()), //double getEnhOrthoScale(isAux()),
-							(imgdtt_params.lma_debug_level > 1)); // boolean debug);
+//							double [][]  corrs = new double [GPUTileProcessor.NUM_PAIRS][corr_length]; // 225-long (15x15)
 							
- */
 							Correlation2d corr2d = new Correlation2d(
 									imgdtt_params,              // ImageDttParameters  imgdtt_params,
 									transform_size,             // int transform_size,
@@ -532,6 +504,7 @@ public class ImageDtt extends ImageDttCPU {
 									(imgdtt_params.lma_debug_level > 1)); // boolean debug);
 
 							for (int indx_tile = ai.getAndIncrement(); indx_tile < num_tiles; indx_tile = ai.getAndIncrement()) {
+								double [][]  corrs = new double [GPUTileProcessor.NUM_PAIRS][corr_length]; // 225-long (15x15)
 								int indx_corr = indx_tile * num_tile_corr;
 								int nt = (corr_indices[indx_corr] >> GPUTileProcessor.CORR_NTILE_SHIFT);
 								int tileX = nt % tilesX;
@@ -543,7 +516,7 @@ public class ImageDtt extends ImageDttCPU {
 									assert pair < GPUTileProcessor.NUM_PAIRS : "invalid correllation pair";
 									pair_mask |= (1 << pair);
 									for (int i = 0; i < corr_length; i++) {
-										corrs[pair][i] = fcorr2D[indx_corr][i]; // from float do double
+										corrs[pair][i] = gpu_corr_scale * fcorr2D[indx_corr][i]; // from float to double
 									}
 									indx_corr++; 
 								}
@@ -934,6 +907,14 @@ public class ImageDtt extends ImageDttCPU {
 													}
 												}
 											}
+										} else { // if (lma != null)
+											if (imgdtt_params.corr_poly_only) { // discard tile if LMA failed
+											    disp_str[0] =  Double.NaN;
+											    disp_str[1] =  0.0;
+											    disparity_map[DISPARITY_INDEX_CM]       [tIndex] = disp_str[0];
+											    disparity_map[DISPARITY_STRENGTH_INDEX] [tIndex] = disp_str[1];
+												
+											}
 										}
 									}
 								} // end of if (corr_stat != null)

src/main/java/com/elphel/imagej/tileprocessor/ImageDttCPU.java

@@ -2440,6 +2440,8 @@ public class ImageDttCPU {
 			debug_offsets[i][j] = imgdtt_params.lma_dbg_offset[i][j]*imgdtt_params.lma_dbg_scale;
 		}
 		
+		final double [] dbg_corr_shift =((imgdtt_params.pcorr_dbg_offsx != 0.0) || (imgdtt_params.pcorr_dbg_offsy != 0.0))?
+				(new double [] {imgdtt_params.pcorr_dbg_offsx,imgdtt_params.pcorr_dbg_offsy}):null;
 
 		final boolean macro_mode = macro_scale != 1;      // correlate tile data instead of the pixel data
 		final int quad = 4;   // number of subcameras
@@ -2568,6 +2570,10 @@ public class ImageDttCPU {
 
 		final double [] filter =     doubleGetCltLpfFd(corr_sigma);
 		
+		final double [] filter_rb =      doubleGetCltLpfFd(imgdtt_params.pcorr_sigma_rb);
+		final double [] filter_common =  doubleGetCltLpfFd(imgdtt_params.getCorrSigma(isMonochrome()));
+		final double pcorr_fat_zero =    imgdtt_params.getFatZero(isMonochrome());
+
 		// prepare disparity maps and weights
 		final int max_search_radius = (int) Math.abs(max_corr_radius); // use negative max_corr_radius for squares instead of circles?
 		final int max_search_radius_poly = 1;
@@ -2672,7 +2678,7 @@ public class ImageDttCPU {
 							}
 						}
 						boolean debugTile =(tileX == debug_tileX) && (tileY == debug_tileY) && (globalDebugLevel > -1);
-
+						boolean debugTile0 =(tileX == debug_tileX) && (tileY == debug_tileY) && (globalDebugLevel > -3);
 						final int [] overexp_all = (saturation_imp != null) ? ( new int [2]): null;
 
 						// Moved from inside chn loop
@@ -3055,6 +3061,59 @@ public class ImageDttCPU {
 						    // Debug feature - only calculated if requested
 //						    if ((clt_corr_partial != null) && imgdtt_params.corr_mode_debug) {
 						    if ((clt_corr_partial != null) && (imgdtt_params.corr_mode_debug || imgdtt_params.gpu_mode_debug)) {
+						    	if (debugTile0) {
+						    		System.out.println("Debugging tileY = "+tileY+"  tileX = " + tileX);
+						    		System.out.println("Debugging tileY = "+tileY+"  tileX = " + tileX);
+						    		System.out.println("Debugging tileY = "+tileY+"  tileX = " + tileX);
+						    	}
+						    	
+						    	double [][][] corr_pairs_td = corr2d.correlateCompositeTD(
+						    			clt_data,            // double [][][][][][] clt_data,
+						    			tileX,               // int                 tileX,
+						    			tileY,               // int                 tileY,
+						        		0x3f,                // int                 pairs_mask,
+						        		filter_rb,           // double []           lpf_rb, // extra lpf for red and blue (unused for mono) or null
+						        		getScaleStrengths(), // double              scale_value, // scale correlation value
+						        		col_weights);        // double []           col_weights);
+
+						    	double [][] corrs_ortho_td = corr2d.combineCorrelationsTD(
+						        		corr_pairs_td,       // double [][][] corr_combo_td, // vertical will be transposed, other diagonal flipped vertically (should be separate)
+						        		0x0f);               // int           pairs_mask); // vertical
+						    	if (dbg_corr_shift!= null) {
+						    		fract_shift(    // fractional shift in transform domain. Currently uses sin/cos - change to tables with 2? rotations // USED in lwir
+						    				corrs_ortho_td, // double  [][]  clt_tile,
+						    				dbg_corr_shift[0], // double        shiftX,
+						    				dbg_corr_shift[1], // double        shiftY,
+						    				false); // boolean       bdebug);
+						    	}
+						    	
+						    	corr2d.normalize_TD(
+						    			corrs_ortho_td,      // double [][] td,
+						    			filter_common,       // double []   lpf, // or null
+						    			pcorr_fat_zero);     // double      fat_zero);
+						    	
+						    	double [] corrs_ortho = corr2d.convertCorrToPD(
+						    			corrs_ortho_td);     // double [][] td);
+
+						    	double [][] corrs_cross_td = corr2d.combineCorrelationsTD(
+						        		corr_pairs_td,       // double [][][] corr_combo_td, // vertical will be transposed, other diagonal flipped vertically (should be separate)
+						        		0x30);               // int           pairs_mask); // vertical
+						    	if (dbg_corr_shift!= null) {
+						    		fract_shift(    // fractional shift in transform domain. Currently uses sin/cos - change to tables with 2? rotations // USED in lwir
+						    				corrs_cross_td, // double  [][]  clt_tile,
+						    				dbg_corr_shift[0], // double        shiftX,
+						    				dbg_corr_shift[1], // double        shiftY,
+						    				false); // boolean       bdebug);
+						    	}
+
+						    	corr2d.normalize_TD(
+						    			corrs_cross_td,      // double [][] td,
+						    			filter_common,       // double []   lpf, // or null
+						    			pcorr_fat_zero);     // double      fat_zero);
+						    	
+						    	double [] corrs_cross = corr2d.convertCorrToPD(
+						    			corrs_cross_td);     // double [][] td);
+						    	
 							    double [] strip_ortho = corr2d.combineInterpolatedCorrelations(
 							    		strips,                         // double [][] strips,
 							    		0x0f,                           // int         pairs_mask,
@@ -3093,8 +3152,10 @@ public class ImageDttCPU {
 						    	clt_corr_partial[tileY][tileX][0][3] = corrs[3];                        // 4
 						    	clt_corr_partial[tileY][tileX][1][0] = corrs[4];                        // 5
 						    	clt_corr_partial[tileY][tileX][1][1] = corrs[5];                        // 6
-						    	clt_corr_partial[tileY][tileX][1][2] = corr2d.debugStrip(strip_hor);    // 7
-						    	clt_corr_partial[tileY][tileX][1][3] = corr2d.debugStrip(strip_vert);   // 8
+						    	clt_corr_partial[tileY][tileX][1][2] = corrs_ortho;                     // 7
+						    	clt_corr_partial[tileY][tileX][1][3] = corrs_cross;                     // 8
+//						    	clt_corr_partial[tileY][tileX][1][2] = corr2d.debugStrip(strip_hor);    // 7
+//						    	clt_corr_partial[tileY][tileX][1][3] = corr2d.debugStrip(strip_vert);   // 8
 						    	clt_corr_partial[tileY][tileX][2][0] = corr2d.debugStrip(strips[4]);    // 9
 						    	clt_corr_partial[tileY][tileX][2][1] = corr2d.debugStrip(strips[5]);    // 10
 						    	clt_corr_partial[tileY][tileX][2][2] = corr2d.debugStrip2(strip_hor);   // 11

src/main/java/com/elphel/imagej/tileprocessor/QuadCLT.java

@@ -277,7 +277,7 @@ public class QuadCLT extends QuadCLTCPU {
 		};
 		gpuQuad.setLpfRbg( // constants memory - same for all cameras
 				lpf_rgb,
-				!batch_mode);
+				true); // !batch_mode);
 
 		float [] lpf_flat = image_dtt.floatGetCltLpfFd(clt_parameters.getGpuCorrSigma(is_mono));
 
@@ -499,10 +499,20 @@ public class QuadCLT extends QuadCLTCPU {
 		boolean is_lwir = quadCLT_main.isLwir();
 
 		double    fat_zero = clt_parameters.getGpuFatZero(is_mono); //   30.0;
+/*		
 		double [] scales = (is_mono) ? (new double [] {1.0}) :(new double [] {
 				clt_parameters.gpu_weight_r, // 0.25
 				clt_parameters.gpu_weight_b, // 0.25
 				1.0 - clt_parameters.gpu_weight_r - clt_parameters.gpu_weight_b}); // 0.5
+*/
+		double corr_green = 1.0/(1.0 + clt_parameters.gpu_weight_r + clt_parameters.gpu_weight_b);    // green color
+		// the same, as col_weights, for separate debugging
+		double [] scales= (is_mono) ? (new double [] {1.0}) :(new double [] {
+				clt_parameters.gpu_weight_r *  corr_green,
+				clt_parameters.gpu_weight_b * corr_green,
+				corr_green});
+		
+		
 		double cwgreen = 1.0/(1.0 + clt_parameters.corr_red + clt_parameters.corr_blue);    // green color
 		double [] col_weights= (is_mono) ? (new double [] {1.0}) :(new double [] {
 				clt_parameters.corr_red *  cwgreen,
@@ -522,20 +532,20 @@ public class QuadCLT extends QuadCLTCPU {
 		};
 		quadCLT_main.getGPU().setLpfRbg( // constants memory - same for all cameras
 				lpf_rgb,
-				debugLevel > -1);
+				debugLevel > -1); // -3
 
 		float [] lpf_flat = image_dtt.floatGetCltLpfFd(clt_parameters.getGpuCorrSigma(is_mono));
 
 		quadCLT_main.getGPU().setLpfCorr(// constants memory - same for all cameras
 				"lpf_corr", // String const_name, // "lpf_corr"
 				lpf_flat,
-				debugLevel > -1);
+				debugLevel > -1); // -3
 
 		float [] lpf_rb_flat = image_dtt.floatGetCltLpfFd(clt_parameters.getGpuCorrRBSigma(is_mono));
 		quadCLT_main.getGPU().setLpfCorr(// constants memory - same for all cameras
 				"lpf_rb_corr", // String const_name, // "lpf_corr"
 				lpf_rb_flat,
-				debugLevel > -1);
+				debugLevel > -1); // -3
 
 		final boolean use_aux = false; // currently GPU is configured for a single quad camera
 
@@ -1171,7 +1181,9 @@ public class QuadCLT extends QuadCLTCPU {
 				  texture_woi,                   // texture_woi,     // null or generated texture location/size
 				  null,                          //  iclt_fimg,       // will return quad images or null to skip, use quadCLT.linearStackToColor 
 					// new parameters, will replace some other?
-				  clt_parameters.getFatZero(isMonochrome()),      // final double              gpu_fat_zero,    // clt_parameters.getGpuFatZero(is_mono);absolute == 30.0
+//				  clt_parameters.getFatZero(isMonochrome()),      // final double              gpu_fat_zero,    // clt_parameters.getGpuFatZero(is_mono);absolute == 30.0\
+				  clt_parameters.gpu_corr_scale, //  gpu_corr_scale,  //  0.75; // reduce GPU-generated correlation values
+				  clt_parameters.getGpuFatZero(isMonochrome()), // final double              gpu_fat_zero,    // clt_parameters.getGpuFatZero(is_mono);absolute == 30.0\
 				  clt_parameters.gpu_sigma_r,    // 0.9, 1.1
 				  clt_parameters.gpu_sigma_b,    // 0.9, 1.1
 				  clt_parameters.gpu_sigma_g,    // 0.6, 0.7

src/main/java/com/elphel/imagej/tileprocessor/QuadCLTCPU.java

@@ -9954,7 +9954,7 @@ public class QuadCLTCPU {
 						  tilesX*(2*image_dtt.transform_size),
 						  tilesY*(2*image_dtt.transform_size),
 						  true,
-						  image_name+sAux()+"-PART_CORR-GPU-D"+clt_parameters.disparity);
+						  image_name+sAux()+"-PART_CORR-CPU-D"+clt_parameters.disparity);
 			  }
 		  }