started TD neighbors

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

@@ -507,7 +507,7 @@ public class GpuQuad{ // quad camera description
 	public int getSensorWidth()  {return this.img_width;}
 	public int getSensorHeight() {return this.img_height;}
 	
-	public int getDttSize()     {return GPUTileProcessor.DTT_SIZE;}
+	public static int getDttSize()     {return GPUTileProcessor.DTT_SIZE;}
 //	public int getNumCams()     {return GPUTileProcessor.NUM_CAMS;}
 	public int getSensorMaskInter() {return sensor_mask_inter;}
 	public void setSensorMaskInter(int mask) {
@@ -2718,7 +2718,6 @@ public class GpuQuad{ // quad camera description
 	public int getNumPairs() {return num_all_pairs;}
 	
 
-	// 
 	/**
 	 * Generating correlation sequence by CPU to correlate all tiles provided in linescan order.
 	 * Additionally, if both (num_acc != null) and (pfcorr_weights !=null), pfcorr_weights[0]
@@ -2817,10 +2816,10 @@ public class GpuQuad{ // quad camera description
 			int ty = tp_tasks[ntile].ty;
 			int tx = tp_tasks[ntile].tx;
 			for (int ipair = 0; ipair < num_pairs; ipair++) {
-				int pair = inter_mode?getSensInter(ipair): getCorrPair(ipair);
+				int pair = inter_mode?getSensInter(ipair): getCorrPair(ipair); // 0xff for accumulated
 				int corr_pair = ntile * num_pairs + ipair;
 				// Below was an error (not visible if all selected, using index of the pair, not an absolute pair )
-				indices[corr_pair]= // ntile * num_pairs + pair] = 
+				indices[corr_pair]= // ntile * num_pairs + pair] = // low 8 bits - pair, 0xff for sum of all channels
 						((ty * tilesX + tx) << GPUTileProcessor.CORR_NTILE_SHIFT) +
 						(pair & ((1 <<  GPUTileProcessor.CORR_NTILE_SHIFT) -1) );
 				if (fcorr_weights != null) {
@@ -2994,9 +2993,44 @@ public class GpuQuad{ // quad camera description
 		return corr_tiles;
 	}
 
+	/**
+	 * Read GPU correlation data to the sparse array [tilesY][tilesX][] for a single correlation pair (usually a combo one)
+	 * @param inter true for interscene correlations, false - for the intrascene ones
+	 * @param pair correlation pair index to read
+	 * @return an array of transform-domain correlations mapped to tiles [tilesY][tilesX]. Each tile data is either null
+	 *         or [4*64] float array
+	 */
+	public float [][][] getCorrTilesLayerTD(
+			boolean        inter,
+			int            pair) {
+		int tilesX =     img_width / GPUTileProcessor.DTT_SIZE;
+		int tilesY =     img_height / GPUTileProcessor.DTT_SIZE;
+        int num_pairs = inter? (getNumCamsInter() + 1 + 0): getNumUsedPairs();// Number of used pairs		num_pairs = num_pairs_in;
+		final int corr_size_td = 4 * GPUTileProcessor.DTT_SIZE * GPUTileProcessor.DTT_SIZE;
+		int [] indices = getCorrIndices(); // also sets num_corr_tiles
+		float [] fdata = getCorrTdData();
+		int num_tiles = num_corr_tiles / num_pairs;
+		float [][][] corr_tiles = new float [tilesY][tilesX][];
+
+		for (int nt = 0; nt < num_tiles; nt++ ) {
+			int nTile = (indices[nt * num_pairs] >> GPUTileProcessor.CORR_NTILE_SHIFT);
+			int ty = nTile / tilesX;
+			int tx = nTile % tilesX;
+
+			corr_tiles[ty][tx] = new float [corr_size_td];
+			System.arraycopy(
+					fdata,
+					(nt * num_pairs + pair) * corr_size_td,
+					corr_tiles[ty][tx],
+					0,
+					corr_size_td);
+		}
+		return corr_tiles;
+	}
+
 
 
-	public int [] setCorrTilesComboTd(
+	public int [] setCorrTilesComboTd( // not used?
 			final float [][][] corr_tiles, // [tileY][tileX][4*64]
 			int ipair) // just to set in the index low bits
 	{
@@ -3025,7 +3059,7 @@ public class GpuQuad{ // quad camera description
 		return indices_trim;
 	}
 
-	public float [][][] getCorrTilesComboTd() // [tileY][tileX][4*64] , read all available pairs
+	public float [][][] getCorrTilesComboTd() // [tileY][tileX][4*64] , read all available pairs  // not used?
 	{
 		int tilesX =     img_width / GPUTileProcessor.DTT_SIZE;
 		int tilesY =     img_height / GPUTileProcessor.DTT_SIZE;

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

@@ -44,6 +44,7 @@ import Jama.Matrix;
 import ij.ImageStack;
 
 public class ImageDttCPU {
+	static final int THREADS_MAX = 100;
 	static boolean FPGA_COMPARE_DATA= false; // true; // false; //
 	static int     FPGA_SHIFT_BITS =  7; // number of bits for fractional pixel shift
 	static int     FPGA_PIXEL_BITS = 15; // bits to represent pixel data (positive)

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

@@ -12998,6 +12998,9 @@ public class OpticalFlow {
 
 		scene.saveQuadClt(); // to re-load new set of Bayer images to the GPU (do nothing for CPU) and Geometry
 		float  [][][][]     fcorr_td =  null; // no accumulation, use data in GPU
+		// Generate 2D phase correlations from the CLT representation
+		// generates sum of the per-channel correlations as the last slot.
+		// updates gpuQuad.gpu_corr_indices, gpuQuad.gpu_corrs_td and some other
 		if (mb_en && (mb_vectors!=null)) {
 			image_dtt.interCorrTDMotionBlur(
 					clt_parameters.img_dtt,     // final ImageDttParameters  imgdtt_params,    // Now just extra correlation parameters, later will include, most others
@@ -13158,10 +13161,12 @@ public class OpticalFlow {
 		double scale_neibs_td = use_neibs? clt_parameters.imp.scale_neibs_td : 0; // scale threshold for the transform-domain average maximums
 		double scale_avg_weight =  clt_parameters.imp.scale_avg_weight;           // reduce influence of the averaged correlations compared to the single-tile ones
 		
+		int [] corr_indices_dbg = show_2d_correlations? image_dtt.getGPU().getCorrIndices() : null;
 		coord_motion = image_dtt.clt_process_tl_interscene(       // convert to pixel domain and process correlations already prepared in fcorr_td and/or fcorr_combo_td
 				clt_parameters.img_dtt,            // final ImageDttParameters  imgdtt_params,   // Now just extra correlation parameters, later will include, most others
 				// only used here to keep extra array element for disparity difference
 				use3D,                             // boolean        use3D,         // generate disparity difference
+				use_neibs,				           // boolean                   use_neibs,
 				fcorr_td,                          // final float  [][][][]     fcorr_td,        // [tilesY][tilesX][pair][4*64] transform domain representation of all selected corr pairs
 				null,                              // float [][][]              num_acc,         // number of accumulated tiles [tilesY][tilesX][pair] (or null). Can be inner null if not used in tp_tasks
 				null,                              // double []                 dcorr_weight,    // alternative to num_acc, compatible with CPU processing (only one non-zero enough)
@@ -13307,9 +13312,12 @@ public class OpticalFlow {
 			float [][][] fclt_corr1 = ImageDtt.convertFcltCorr( // partial length, matching corr_indices = gpuQuad.getCorrIndices(); // also sets num_corr_tiles
 					dcorr_tiles, // double [][][] dcorr_tiles,// [tile][sparse, correlation pair][(2*transform_size-1)*(2*transform_size-1)] // if null - will not calculate
 					fclt_corr);  // float  [][][] fclt_corr) //  new float [tilesX * tilesY][][] or null
+			if (use_neibs) {
+				
+			}
 			float [][] dbg_corr_rslt_partial = ImageDtt.corr_partial_dbg( // not used in lwir
 					fclt_corr1, // final float  [][][]     fcorr_data,       // [tile][pair][(2*transform_size-1)*(2*transform_size-1)] // if null - will not calculate
-					image_dtt.getGPU().getCorrIndices(), // tp_tasks,  // final TpTask []         tp_tasks,        //
+					corr_indices_dbg, // image_dtt.getGPU().getCorrIndices(), // tp_tasks,  // final TpTask []         tp_tasks,        //
 					tilesX,    //final int                tilesX,
 					tilesY,    //final int                tilesX,
 					2*image_dtt.transform_size - 1,	// final int               corr_size,